Datasets:

jamescalam

/

ai-arxiv-chunked

like 32

2209.07686

263

Preceding sections empirically underscore the prominence of text and patterns and their intertwined pathways to generate correct answers. We articulate our findings in the following. Patterns are the hidden force that help generate meaningful text. Without patterns the model is not purposefully prompted to generate intermediate meaningful text. In addition, the presence of patterns reinforces how the model should form factual connections between different clauses in the intermediate text. This intermediate text, which generally embody relevant information to the underlying task, drives the model to elicit knowledge and attain correct conclusions. Text imbues patterns with knowledge and meaning. Theoretically, patterns seem to be sufficient in directing the model to solve a task. However, patterns need explicit text to be effective in imparting the requisite information to the model about how to accomplish a task. The role of text primarily is to elaborate the patterns and imbues them with knowledge and meaning. The presence of text enables the model to elicit requisite information from thoughts and reach to a factual conclusion. We empirically observe the crucial role of the text in assisting the model to solve a task, such as GSM-8 K(e.g., four-legged vs. two-legged animals), S PORTS (e.g., sports that is being talked about), and D ATE(e.g., Christmas Eve is on 24th).

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

264

animals), S PORTS (e.g., sports that is being talked about), and D ATE(e.g., Christmas Eve is on 24th). J. A DDITIONAL RESULTS CCOT: C ONCISE CHAIN OFTHOUGHT J.1. CC OT P ROMPTS J.2. R EDUCTION IN LENGTH BY CCOT In this work, we empirically show how the symbiosis between text and patterns unlocks the functional potential of C OT. Relying on our preceding findings, this section explores the idea of engineering effective prompts tailored to remove ineffectual tokens. The benefit of such tailored design is multi-fold: 1.Reducing noise. Reducing noise from prompts could potentially reduce the confusion for the model and consequently lead to higher task solve rate. Investigating the impact of noise (futile or misleading information) in prompts on the model performance is an interesting future work. 2.Potential energy savings. Last but not the least, reduced number of tokens in the prompts instruct the model to generate less number of tokens per question. Less number of generated tokens directly translate to reduced runtime and carbon footprint per inference. This is especially crucial because of the inherently large compute cost associated with the large language models.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

265

to reduced runtime and carbon footprint per inference. This is especially crucial because of the inherently large compute cost associated with the large language models. Our proposed prompts, dubbed as CC OT9, distill the essence of our hypothesis by supplying minimal requisite information to drive the model to factual conclusions. Table 8 compares C OT and CC OT across different tasks across the four models (PaLM-62 Band PaLM-540 B). In all tasks, except GSM-8 K, we use identical examples in the prompt and only rephrased the thoughts to use less number of tokens. For GSM-8 K, we could not find a systematic mechanism to shorten the thoughts. Instead, we randomly harvest questions from training set whose thoughts are shorter compared to C OT. Overall, CC OT outperform C OT while in general employing prompts with less number of tokens. The task solve rate of CC OT remains relatively high as we scale the model to PaLM-540 B, highlighting the efficiency of CC OT. 9Concise C hain O f Thought 63 Table 33: CC OT prompt for GSM-8 K. qPrompt 1 åQÞÑWhat isfifteen more than aquarterof48?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

266

Table 33: CC OT prompt for GSM-8 K. qPrompt 1 åQÞÑWhat isfifteen more than aquarterof48? TÞÑAquarterof48is48/4=12.15more than 12is12+15=27. VAÞÑ27 qPrompt 2 åQÞÑTwice Angie’s age, plus 4,is20.How oldisAngie? TÞÑAngie’s ageminus4is20-4=16.16istwice Angie’s age. Thus, Angie’s ageis16/2=8. VAÞÑ8 qPrompt 3 åQÞÑSteve is5’6”. Hegrows 6inches. How tallisheininches? TÞÑOne feethas12inches. Initially, Steve is5*12+6=66inches tall.Aftergrow ing6inches, Steve is66+6 =72inches tall. VAÞÑ72 qPrompt 4 åQÞÑ198passengers fitinto9buses. How many passengers fitin5buses?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

267

qPrompt 4 åQÞÑ198passengers fitinto9buses. How many passengers fitin5buses? TÞÑCapacityofonebusis198passengers /9buses =22passengers inonebus. Thus, 5buses canfit22*5= 110passengers. VAÞÑ110 qPrompt 5 åQÞÑFifteen more than aquarterofanumberis27.What isthenumber? TÞÑFifteen lessthan 27is27-15=12.Thequarterofthenumberisthus 12,andthenumberis12*4=48. VAÞÑ48 qPrompt 6 åQÞÑIf12bags oforanges weigh 24pounds, how much do8bags weigh? TÞÑ12bags weight 24pounds, soonebagweights 24pounds /12=2pounds. So8bags oforanges would weigh 8*2pounds =16pounds. VAÞÑ16 qPrompt 7 åQÞÑSandraeats3beignets everymorn ing.How many beignets willsheeatin16weeks?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

268

qPrompt 7 åQÞÑSandraeats3beignets everymorn ing.How many beignets willsheeatin16weeks? TÞÑSandraeats3beignets everymorn ingandthere are7days inaweek sosheeats3*7=21beignets ina week. Sandraeats21beignets inaweek, soin16weeks shewilleat21*16=336beignets. VAÞÑ336 qPrompt 8 åQÞÑJohn takes apillevery6hours. How many pills does hetake aweek? TÞÑThere are24hours inaday. SoJohn takes 24/6=4pills aday. Inaweek, John there fore takes 4*7=28 pills. VAÞÑ28 64 Table 34: CC OT prompt for D ATE. qPrompt 1 åQÞÑItis4/19/1969 today. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayis04/19/1969. 24hours (one day) later is04/20/1969. VAÞÑ04/20/1969 qPrompt 2

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

269

VAÞÑ04/20/1969 qPrompt 2 åQÞÑTheconcertwassched uled tobeon06/01/1943, butwasdelayed byonedaytotoday. What isthedate 10 days agoinMM/D D/YYYY? TÞÑTodayis06/02/1943 (one dayafter06/01/1943). 10days beforetodayis05/23/1943. VAÞÑ05/23/1943 qPrompt 3 åQÞÑThefirstdayof2019 isaTues day, andtodayisthefirstMon dayof2019. What isthedate todayinMM/D D/YYYY? TÞÑ01/01/2019 wasaTues day(first dayof2019). TodayisthefirstMon day, 01/07/2019. (sixdays later). VAÞÑ01/07/2019 qPrompt 4 åQÞÑJane was born onthelastdayofFebru aryin2001. Todayisher16-year-oldbirth day. What isthedate yesterdayinMM/D D/YYYY?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

270

yesterdayinMM/D D/YYYY? TÞÑJane wasborn on02/28/2001. Sotodayis02/28/2017 andyesterdaywas02/27/2017. VAÞÑ02/27/2017 qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑ2015 iscomingin2days (36hours). Sotodayis12/30/2021, andoneweek from todaywillbe01/05/2015. VAÞÑ01/05/2015 qPrompt 6 åQÞÑJane thought todayis3/11/2002, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayis03/12/2002. Sothedate 24hours later willbe03/13/2002. VAÞÑ03/13/2002 65 Table 35: CC OT prompt for S PORTS . qPrompt 1 åQÞÑIsthefollowingsentence plausible? “Jamel Murraywasperfectfrom theline.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

271

qPrompt 1 åQÞÑIsthefollowingsentence plausible? “Jamel Murraywasperfectfrom theline.” TÞÑJamalMurray->basketball. perfectfrom theline->basketball. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑJoao Moutinho ->soccer.NFC cham pionship ->Amer icanfootball. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑJonas Valan ciunas->basketball. beatingthebuzzer ->basketball. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑSam Darnold ->Amer icanfootball. passingthepuck ->hockey. VAÞÑno qPrompt 5

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

272

TÞÑSam Darnold ->Amer icanfootball. passingthepuck ->hockey. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑKyle Palmieri ->hockey. called forslash ing->hockey. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll.” TÞÑCarsonWentz is->Amer icanfootball. pick androll->basketball. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑMalcolm Brog don->basketball. bank ingtheshot in->basketball. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑDray mond Green ->basketball. throw ingatouch down ->football. VAÞÑno

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

273

TÞÑDray mond Green ->basketball. throw ingatouch down ->football. VAÞÑno 66 Table 36: Comparison of the average number of input and output tokens for PaLM-62 B. The corresponding tables for each task refers to C OT and CC OT prompts, respectively. TaskAvg. # of Input Tokens Avg. # of Output Tokens COT CC OT Reduction C OT CC OT Reduction GSM-8 K(COTÞÑTable 38, CC OTÞÑTable 33) 503.25 434.25 1.16  55.51 62.02 0.90  DATE(COTÞÑTable 39, CC OTÞÑTable 34) 321.39 276.39 1.16  21.47 15.55 1.38  SPORTS (COTÞÑTable 40, CC OTÞÑTable 35) 283.97 240.97 1.18  20.04 15.27 1.31  SORTING (COTÞÑTable 86, CC OTÞÑTable 41) 553.01 266.01 2.08  99.34 36.16 2.75  Average Reduction — — 1.39 — — 1.58

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

274

Average Reduction — — 1.39 — — 1.58 The next logical question is to determine to what extend CC OT reduces the number of tokens. Table 36 compares the average number of input/output tokens between C OT and CC OT. On average, CC OT (our approach) reduces the number of input and output tokens by 1.39 and 1.58, respectively. This reduction in the number of tokens could potentially render as energy saving in the model inference, which we defer for future work. To conclude, we highlight the importance of conducting theoretical and empirical studies to better understand the sensitivity of LLM performance to the number of tokens under few-shot setting. We now present additional results and examples that were moved from the main text for brevity and clarity. Table 37: Some representative examples from small number of cases where just pattern is better åQ1ÞÑIs the following sentence plausible? "Gardner Minshew was safe at home." A(p): Gardner Minshew is a football player. Being safe at home is part of life. So the answer is yes. A(Cpat_noneppq): Both are a part of different sports. So the answer is no

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

275

A(Cpat_noneppq): Both are a part of different sports. So the answer is no åQ2ÞÑBrandon’s iPhone is four times as old as Ben’s iPhone. Ben’s iPhone is two times older than Suzy’s iPhone. If Suzy’s iPhone is 1 year old, how old is Brandon’s iPhone? A(p): Brandon’s iPhone is 4 times as old as Ben’s iPhone. So Brandon’s iPhone is 4 years old. Ben’s iPhone is 2 times older than Suzy s iPhone. So Ben s iPhone is 2 years older than Suzy s iPhone. So Suzy’s iPhone is 1 year old. The answer is 4. A(Cpat_noneppq): 4 * 2 * 1 = 8. The answer is 8. åQ3ÞÑJerry has two pools, both with leaks emptying them out at 4 gallons/minute. 4 minutes ago, the big pool had twice as much water as the small pool. Now the big pool has four times as much water as the small pool. How much water does the small pool have now? A(p): The big pool had twice as much water as the small pool. So the big pool had 2 * 4 = 8 gallons. The small pool

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

276

A(p): The big pool had twice as much water as the small pool. So the big pool had 2 * 4 = 8 gallons. The small pool had 4 gallons. 4 minutes ago, the big pool had 8 gallons and the small pool had 4 gallons. Now the big pool has 4 times as much water as the small pool. So the big pool has 4 * 4 = 16 gallons. The small pool has 4 - 16 = -12 gallons. The answer is -12. A(Cpat_noneppq): 2 * 4 = 8. The answer is 8. 67 Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango Source Code 1: Python code to reproduce Sort dataset. 1 from itertools import combinations 2 from itertools import permutations 3 import numpy asnp 4 5 def GenerateCombinations (A, r): 6 """ 7 Returns subsequences of elements with length r from array A. 8 """ 9 return list (combinations(A, r)) 10 11 def main (_): 12 base_array =np.arange( 10) 13 with open (sort_dataset.tsv ,wt)assort_ds: 14 tsv_writer =csv.writer(sort_ds, delimiter =\t)

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

277

14 tsv_writer =csv.writer(sort_ds, delimiter =\t) 15 tsv_writer .writerow([ Question ,Answer ,Thought ]) 16 for jinrange (2,10): 17 current_array =GenerateCombinations(base_array, j) 18 for iincurrent_array: 19 current_permutations =list (permutations(i)) 20 for sincurrent_permutations: 21 question =,.join([ str(x) for xins]) 22 answer =,.join([ str(x) for xinnp.sort( list (s))]) 23 thought =<.join([ str(x) for xinnp.sort( list (s))]) 24 tsv_writer .writerow([question, answer, thought]) I.COMPLETE LIST OF COUNTERFACTUAL PROMPTS This section includes all the counterfactual prompts used in our experiments. 59Figure 16: Python code to reproduce Sort dataset. K. C OMPLETE LIST OF COUNTERFACTUAL PROMPTS This section includes all the counterfactual prompts used in our experiments. 68 Table 38: GSM-8 Kprompts used by Wei et al. (2022).

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

278

This section includes all the counterfactual prompts used in our experiments. 68 Table 38: GSM-8 Kprompts used by Wei et al. (2022). qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does he have now? TÞÑShawn started with 5toys. Ifhegot2toys each from hismom anddad, then thatis4more toys. 5+4=9. VAÞÑ9 qPrompt 2 åQÞÑIfthere are3carsintheparkinglotand2more carsarrive, how many carsareintheparkinglot? TÞÑThere areoriginally 3cars. 2more carsarrive. 3+2=5. VAÞÑ5 qPrompt 3 åQÞÑJasonhad20lollipops. Hegave Denny some lollipops. Now Jasonhas12lollipops. How many lollipops did Jasongive toDenny?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

279

Jasongive toDenny? TÞÑJasonstarted with 20lollipops. Then hehad12aftergivingsome toDenny. Sohegave Denny 20-12=8. VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from Mon day toThurs day. How many computersarenow intheserver room? TÞÑThere were originally 9computers.Foreach of4days, 5more computerswere added. So5*4=20 computerswere added. 9+20is29. VAÞÑ29 qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbe21trees. How many trees didthegrove work ersplant today? TÞÑThere are15trees originally. Then there were 21trees aftersome more were planted. Sothere must have been 21-15=6. VAÞÑ6 qPrompt 6

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

280

been 21-15=6. VAÞÑ6 qPrompt 6 åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate35,how many pieces dothey have leftintotal? TÞÑOriginally, Leah had32choco lates. Hersisterhad42.Sointotalthey had32+42=74.Aftereating35,they had74-35=39. VAÞÑ39 qPrompt 7 åQÞÑOlivia has$23. Shebought fivebagels for$3each. How much money does shehave left? TÞÑOlivia had23dollars. 5bagels for3dollarseach willbe5x3=15dollars. Soshehas23-15dollarsleft.23 -15is8. VAÞÑ8 qPrompt 8 åQÞÑMichael had58golfballs. OnTues day, helost23golfballs. OnWednes day, helost2more. How many golf balls didhehave attheendofWednes day?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

281

balls didhehave attheendofWednes day? TÞÑMichael started with 58golfballs. Afterlosing23onTues day, hehad58-23=35.Afterlosing2more, he had35-2=33golfballs. VAÞÑ33 69 Table 39: D ATEprompts used by Wei et al. (2022). qPrompt 1 åQÞÑItis4/19/1969 today. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayis04/19/1969. 24hours later isonedayaftertoday, which would be04/20/1969. VAÞÑ04/20/1969 qPrompt 2 åQÞÑTheconcertwassched uled tobeon06/01/1943, butwasdelayed byonedaytotoday. What isthedate 10 days agoinMM/D D/YYYY? TÞÑOne dayafter06/01/1943 is06/02/1943, sotodayis06/02/1943. 10days beforetodayis05/23/1943. VAÞÑ05/23/1943 qPrompt 3

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

282

VAÞÑ05/23/1943 qPrompt 3 åQÞÑThefirstdayof2019 isaTues day, andtodayisthefirstMon dayof2019. What isthedate todayinMM/D D/YYYY? TÞÑIfthefirstdayof2019 wasTues day, then 01/01/2019 wasaTues day. TodayisthefirstMon day, would be sixdays later. Sotodayis01/07/2019. VAÞÑ01/07/2019 qPrompt 4 åQÞÑJane was born onthelastdayofFebru aryin2001. Todayisher16-year-oldbirth day. What isthedate yesterdayinMM/D D/YYYY? TÞÑThelastdayofFebru aryisthe28th, soJane wasborn on02/28/2001. Todayisher16-year oldbirth day, so todayis02/28/2017. Soyesterdaywas02/27/2017. VAÞÑ02/27/2017 qPrompt 5

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

283

todayis02/28/2017. Soyesterdaywas02/27/2017. VAÞÑ02/27/2017 qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑIf2015 iscomingin36hours, then itiscomingin2days. 2days before01/01/2015 is12/30/2014, sotoday is12/30/2014. Sooneweek from todaywillbe01/05/2015. VAÞÑ01/05/2015 qPrompt 6 åQÞÑJane thought todayis3/11/2002, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayis03/12/2002. Sothedate 24hours later willbe03/13/2002. VAÞÑ03/13/2002 70 Table 40: S PORTS prompts used by Wei et al. (2022). qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

284

qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.” TÞÑJamalMurrayisabasketballplayer. Beingperfectfrom thelineispartofbasketball. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑJoao Moutinho isasoccerplayer. TheNFC cham pionship ispartofAmer icanfootball, notsoccer. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑJonas Valan ciunasisabasketballplayer. Beatingthebuzzer ispartofbasketball. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

285

VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑSam Darnold isaAmer icanfootballplayer. Passingthepuck ispartofhockey, notAmer icanfootball. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑKyle Palmieri isahockey player. Beingcalled forslash ingispartofhockey. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll.” TÞÑCarsonWentz isanAmer icanfootballplayer. Pick androllispartofbasketball, notfootball. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑMalcolm Brog donisabasketballplayer. Bank ingtheshot inispartofbasketball. VAÞÑyes

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

286

TÞÑMalcolm Brog donisabasketballplayer. Bank ingtheshot inispartofbasketball. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑDray mond Green isanbasketballplayer. Throw ingatouch down ispartoffootball, notbasketball. VAÞÑno 71 Table 41: S ORTING prompts using code snippet 16. qPrompt 1 åQÞÑ7,8,4,1,2,9,3,6,5 TÞÑ1<2<3<4<5<6<7<8<9 VAÞÑ1,2,3,4,5,6,7,8,9 qPrompt 2 åQÞÑ5,9,3,1,8,4,6,2 TÞÑ1<2<3<4<5<6<8<9 VAÞÑ1,2,3,4,5,6,8,9 qPrompt 3 åQÞÑ6,5,7,4,3,2,8,1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

287

qPrompt 3 åQÞÑ6,5,7,4,3,2,8,1 TÞÑ1<2<3<4<5<6<7<8 VAÞÑ1,2,3,4,5,6,7,8 qPrompt 4 åQÞÑ1,6,4,8,5,3,7,2 TÞÑ1<2<3<4<5<6<7<8 VAÞÑ1,2,3,4,5,6,7,8 qPrompt 5 åQÞÑ5,2,1,4,3,7 TÞÑ1<2<3<4<5<7 VAÞÑ1,2,3,4,5,7 qPrompt 6 åQÞÑ3,8,2,5,6,4,7,1 TÞÑ1<2<3<4<5<6<7<8 VAÞÑ1,2,3,4,5,6,7,8 qPrompt 7 åQÞÑ8,6,1,2,9,7,4

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

288

qPrompt 7 åQÞÑ8,6,1,2,9,7,4 TÞÑ1<2<4<6<7<8<9 VAÞÑ1,2,4,6,7,8,9 qPrompt 8 åQÞÑ7,6,8,1 TÞÑ1<6<7<8 VAÞÑ1,6,7,8 72 Table 42: D ATEwith abstract dates. qPrompt 1 åQÞÑItisDATE today. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayisDATE. 24hours later isonedayaftertoday, which would beDATE. VAÞÑ04/20/1969 qPrompt 2 åQÞÑTheconcertwassched uled tobeonDATE, butwasdelayed byonedaytotoday. What isthedate 10days agoinMM/D D/YYYY? TÞÑOne dayafterDATE isDATE, sotodayisDATE. 10days beforetodayis05/23/1943. VAÞÑ05/23/1943

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

289

VAÞÑ05/23/1943 qPrompt 3 åQÞÑThefirstdayof2019 isaTues day, andtodayisthefirstMon dayof2019. What isthedate todayinMM/D D/YYYY? TÞÑIfthefirstdayof2019 wasTues day,then DATE wasaTues day.TodayisthefirstMon day,would besixdays later. Sotodayis01/07/2019. VAÞÑ01/07/2019 qPrompt 4 åQÞÑJane was born onthelastdayofFebru aryin2001. Todayisher16-year-oldbirth day. What isthedate yesterdayinMM/D D/YYYY? TÞÑThelastdayofFebru aryisthe28th, soJane wasborn onDATE. Todayisher16-year oldbirth day, sotoday isDATE. Soyesterdaywas02/27/2017. VAÞÑ02/27/2017 qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

290

qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑIf2015 iscomingin36hours, then itiscomingin2days. 2days beforeDATE isDATE, sotodayisDATE. Sooneweek from todaywillbe01/05/2015. VAÞÑ01/05/2015 qPrompt 6 åQÞÑJane thought todayisDATE, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayisDATE. Sothedate 24hours later willbe03/13/2002. VAÞÑ03/13/2002 73 Table 43: GSM-8 Kafter replacing numbers with Greek letters. qPrompt 1 åQÞÑShawn hastoys. ForChrist mas, hegottoys each from hismom anddad. How many toys does hehave now? TÞÑShawn started withtoys. Ifhegottoys each from hismom anddad, then thatismore toys.+=.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

291

TÞÑShawn started withtoys. Ifhegottoys each from hismom anddad, then thatismore toys.+=. VAÞÑ qPrompt 2 åQÞÑIfthere arecarsintheparkinglotandmore carsarrive, how many carsareintheparkinglot? TÞÑThere areoriginallycars.more carsarrive.+=. VAÞÑ qPrompt 3 åQÞÑJasonhadlollipops. Hegave Denny some lollipops. Now Jasonhaslollipops. How many lollipops did Jasongive toDenny? TÞÑJasonstarted withlollipops. Then hehadaftergivingsome toDenny. Sohegave Denny-=. VAÞÑ qPrompt 4 åQÞÑThere werecomputersintheserver room.more computerswere installed each day, from Mon dayto Thurs day. How many computersarenow intheserver room?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

292

Thurs day. How many computersarenow intheserver room? TÞÑThere were originallycomputers.Foreach offour days,more computerswere added. So*four = computerswere added.+is. VAÞÑ qPrompt 5 åQÞÑThere aretrees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbetrees. How many trees didthegrove work ersplant today? TÞÑThere aretrees originally. Then there weretrees aftersome more were planted. Sothere must have been -=. VAÞÑ qPrompt 6 åQÞÑLeah hadchoco lates andhersisterhad.Ifthey ate,how many pieces dothey have leftintotal? TÞÑOriginally, Leah hadchoco lates. Hersisterhad.Sointotalthey had+=.Aftereating,they had -=. VAÞÑ qPrompt 7

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

293

-=. VAÞÑ qPrompt 7 åQÞÑOlivia has.Shebought fivebagels foreach. How much money does shehave left? TÞÑOlivia haddollars. 5bagels fordollarseach willbe5x=dollars. Soshehas-dollarsleft.- is. VAÞÑ qPrompt 8 åQÞÑMichael hadgolf balls. OnTues day, helostgolf balls. OnWednes day, helostmore. How many golf balls didhehave attheendofWednes day? TÞÑMichael started withgolfballs. AfterlosingonTues day, hehad-=.Afterlosingmore, hehad -=golfballs. VAÞÑ 74 Table 44: S ORTING after replacing numbers with Greek letters. qPrompt 1 åQÞÑ,,,,, ,c,, TÞÑc<< <<<<<<

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

294

TÞÑc<< <<<<<< VAÞÑc,, ,,,,,, qPrompt 2 åQÞÑ,,,,, , , TÞÑ<<<< < << VAÞÑ,,,, , ,, qPrompt 3 åQÞÑ,,,,,,, TÞÑ<<<<<<< VAÞÑ,,,,,,, qPrompt 4 åQÞÑ,,,,,,,o TÞÑ<<<<<<o< VAÞÑ,,,,,,o, qPrompt 5 åQÞÑ!,o,,,, TÞÑ!<o<<<<

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

295

qPrompt 5 åQÞÑ!,o,,,, TÞÑ!<o<<<< VAÞÑ!,o,,,, qPrompt 6 åQÞÑ,,,,,c,, TÞÑc<< <<<<< VAÞÑc,, ,,,,, qPrompt 7 åQÞÑ, ,,,,c, TÞÑ<<<<< <c VAÞÑ,,,,, ,c qPrompt 8 åQÞÑ,,, TÞÑ<<< VAÞÑ,,, 75 Table 45: S PORTS with abstract sports person. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “PER SON wasperfectfrom theline.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

296

qPrompt 1 åQÞÑIsthefollowingsentence plausible? “PER SON wasperfectfrom theline.” TÞÑPERSON isabasketballplayer. Beingperfectfrom thelineispartofbasketball. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “PER SON caught thescreen pass intheNFC cham pionship.” TÞÑPERSON isasoccerplayer. TheNFC cham pionship ispartofAmer icanfootball, notsoccer. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “PER SON beat thebuzzer.” TÞÑPERSON isabasketballplayer. Beatingthebuzzer ispartofbasketball. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “PER SON passed thepuck.” TÞÑPERSON isaAmer icanfootballplayer. Passingthepuck ispartofhockey, notAmer icanfootball. VAÞÑno qPrompt 5

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

297

VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “PER SON wascalled forslash ing.” TÞÑPERSON isahockey player. Beingcalled forslash ingispartofhockey. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “PER SON setthepick androll.” TÞÑPERSON isanAmer icanfootballplayer. Pick androllispartofbasketball, notfootball. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “PER SON banked theshot in.” TÞÑPERSON isabasketballplayer. Bank ingtheshot inispartofbasketball. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “PER SON threw atouch down.” TÞÑPERSON isanbasketballplayer. Throw ingatouch down ispartoffootball, notbasketball. VAÞÑno 76 Table 46: S PORTS with abstract sports activity. qPrompt 1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

298

VAÞÑno 76 Table 46: S PORTS with abstract sports activity. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “Jamel MurraywasACTIVITY .” TÞÑJamalMurrayisabasketballplayer. BeingACTIVITY ispartofbasketball. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught ACTIVITY .” TÞÑJoao Moutinho isasoccerplayer. TheACTIVITY ispartofAmer icanfootball, notsoccer. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasACTIVITY .” TÞÑJonas Valan ciunasisabasketballplayer. ACTIVITY ispartofbasketball. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold ACTIVITY .”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

299

VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold ACTIVITY .” TÞÑSam Darnold isaAmer icanfootballplayer. ACTIVITY ispartofhockey, notAmer icanfootball. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wasACTIVITY .” TÞÑKyle Palmieri isahockey player. BeingACTIVITY ispartofhockey. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz settheACTIVITY .” TÞÑCarsonWentz isanAmer icanfootballplayer. ACTIVITY ispartofbasketball, notfootball. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donACTIVITY .” TÞÑMalcolm Brog donisabasketballplayer. ACTIVITY ispartofbasketball. VAÞÑyes qPrompt 8

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

300

TÞÑMalcolm Brog donisabasketballplayer. ACTIVITY ispartofbasketball. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green ACTIVITY .” TÞÑDray mond Green isanbasketballplayer. ACTIVITY ispartoffootball, notbasketball. VAÞÑno 77 Table 47: S PORTS with abstract sport. qPrompt 1 åQÞÑIsthefollowingsentence plausible? "Jamel Murraywasperfectfrom theline." TÞÑJamalMurrayisaSPORT1 player. Beingperfectfrom thelineispartofSPORT1. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? "Joao Moutinho caught thescreen pass intheNFC cham pionship." TÞÑJoao Moutinho isaSPORT2 player. TheNFC cham pionship ispartofSPORT3, notSPORT2. VAÞÑno qPrompt 3

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

301

VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? "Jonas Valan ciunasbeat thebuzzer." TÞÑJonas Valan ciunasisaSPORT1 player. Beatingthebuzzer ispartofSPORT1. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? "Sam Darnold passed thepuck." TÞÑSam Darnold isaSPORT3 player. Passingthepuck ispartofSPORT4, notSPORT3. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? "Kyle Palmieri wascalled forslash ing." TÞÑKyle Palmieri isaSPORT4 player. Beingcalled forslash ingispartofSPORT4. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? "CarsonWentz setthepick androll." TÞÑCarsonWentz isanSPORT3 player. Pick androllispartofSPORT1, notSPORT3. VAÞÑno

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

302

TÞÑCarsonWentz isanSPORT3 player. Pick androllispartofSPORT1, notSPORT3. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? "Mal colm Brog donbanked theshot in." TÞÑMalcolm Brog donisaSPORT1 player. Bank ingtheshot inispartofSPORT1. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? "Dray mond Green threw atouch down." TÞÑDray mond Green isanSPORT1 player. Throw ingatouch down ispartofSPORT3, notSPORT1. VAÞÑno 78 Table 48: GSM-8 Kwith decimal numbers. qPrompt 1 åQÞÑShawn hasfiveandahalftoys. ForChrist mas, hegottwoandahalftoys each from hismom anddad. How many toys does hehave now? TÞÑShawn started with 5.5toys. Ifhegot2.5toys each from hismom anddad, then thatis5more toys. 5.5+5

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

303

=10.5. VAÞÑ10.5 qPrompt 2 åQÞÑIfthere are3.3carsintheparkinglotand2.8more carsarrive, how many carsareintheparkinglot? TÞÑThere areoriginally 3.3cars. 2.8more carsarrive. 3.3+2.8=6.1. VAÞÑ6.1 qPrompt 3 åQÞÑJasonhad20.2 lollipops. Hegave Denny some lollipops. Now Jasonhas15.5 lollipops. How many lollipops didJasongive toDenny? TÞÑJasonstarted with 20.2 lollipops. Then hehad15.5 aftergivingsome toDenny. Sohegave Denny 20.2 15.5 =4.7. VAÞÑ4.7 qPrompt 4 åQÞÑThere were nine andaquartercomputersintheserver room. Five andthree quartersmore computerswere installed each day, from Mon daytoThurs day. How many computersarenow intheserver room?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

304

installed each day, from Mon daytoThurs day. How many computersarenow intheserver room? TÞÑThere were originally 9.25 computers.Foreach of4days, 5.75 more computerswere added. So5.75 *4= 23computerswere added. 9.25 +23is32.25. VAÞÑ32.25 qPrompt 5 åQÞÑThere are15.3 trees inthegrove. Grove work erswillplant trees inthegrove today.Afterthey aredone, there willbe16.5 trees. How many trees didthegrove work ersplant today? TÞÑThere are15.3 trees originally. Then there were 16.5 trees aftersome more were planted. Sothere must have been 16.5 -15.3 =1.2. VAÞÑ1.2 qPrompt 6 åQÞÑLeah had3.2choco lates andhersisterhad4.2.Ifthey ate3.5,how many pieces dothey have leftintotal?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

305

TÞÑOriginally, Leah had3.2choco lates. Hersisterhad4.2.Sointotalthey had3.2+4.2=7.4.Aftereating3.5, they had7.4-3.5=3.9. VAÞÑ3.9 qPrompt 7 åQÞÑOlivia has$2.3. Shebought fivebagels for$0.3 each. How much money does shehave left? TÞÑOlivia had2.3dollars. 5bagels for.3dollarseach willbe5x.3=1.5dollars. Soshehas2.3-1.5dollars left.2.3-1.5is0.8. VAÞÑ0.8 qPrompt 8 åQÞÑMichael had5.8golf balls. OnTues day, helost2.3golf balls. OnWednes day, helost0.2more. How many golfballs didhehave attheendofWednes day?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

306

golfballs didhehave attheendofWednes day? TÞÑMichael started with 5.8golfballs. Afterlosing2.3onTues day,hehad5.8-2.3=3.5.Afterlosing0.2more, hehad3.5-0.2=3.3golfballs. VAÞÑ3.3 79 Table 49: D ATEwith future dates. qPrompt 1 åQÞÑItis4/30/3069 today. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayis04/30/3069. 24hours later isonedayaftertoday, which would be04/31/3069. VAÞÑ04/31/3069 qPrompt 2 åQÞÑTheconcertwassched uled tobeon06/01/3043, butwasdelayed byonedaytotoday. What isthedate 10 days agoinMM/D D/YYYY? TÞÑOne dayafter06/01/3043 is06/02/3043, sotodayis06/02/3043. 10days beforetodayis05/23/3043.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

307

VAÞÑ05/23/3043 qPrompt 3 åQÞÑThefirstdayof3130 isaTues day, andtodayisthefirstMon dayof3130. What isthedate todayinMM/D D/YYYY? TÞÑIfthefirstdayof3130 wasTues day, then 01/01/3130 wasaTues day. TodayisthefirstMon day, would be sixdays later. Sotodayis01/07/3130. VAÞÑ01/07/3130 qPrompt 4 åQÞÑJane was born onthelastdayofFebru aryin3101. Todayisher16-year-oldbirth day. What isthedate yesterdayinMM/D D/YYYY? TÞÑThelastdayofFebru aryisthe28th, soJane wasborn on02/28/3101. Todayisher16-year oldbirth day, so todayis02/28/3117. Soyesterdaywas02/27/3117. VAÞÑ02/27/3117 qPrompt 5

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

308

todayis02/28/3117. Soyesterdaywas02/27/3117. VAÞÑ02/27/3117 qPrompt 5 åQÞÑ3115 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑIf3115 iscomingin36hours, then itiscomingin2days. 2days before01/01/3115 is12/30/3114, sotoday is12/30/3114. Sooneweek from todaywillbe01/05/3115. VAÞÑ01/05/3115 qPrompt 6 åQÞÑJane thought todayis3/11/3102, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayis03/12/3102. Sothedate 24hours later willbe03/13/3102. VAÞÑ03/13/3102 80 Table 50: S PORTS with non-sport names and activities. qPrompt 1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

309

VAÞÑ03/13/3102 80 Table 50: S PORTS with non-sport names and activities. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “Adair Fosterwasjuggling thepapercups.” TÞÑAdair Fosterisabasketballplayer. Juggling thepapercups ispartofbasketball. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Scot tieColby caught thehotpotato intheNFC cham pionship.” TÞÑScottieColby isasoccerplayer. TheNFC cham pionship ispartofAmer icanfootball, notsoccer. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Chas Jase beat thebuzzer.” TÞÑChas Jase isabasketballplayer. Beatingthepillowispartofbasketball. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Denny Dillanpassed thesoda.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

310

qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Denny Dillanpassed thesoda.” TÞÑDenny DillanisaAmer icanfootballplayer. Passingthesoda ispartofhockey, notAmer icanfootball. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “MyronStorm wascalled fortrash ing.” TÞÑMyronStorm isahockey player. Beingcalled fortrash ingispartofhockey. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Cliff Kristo pher setthetickandfloor.” TÞÑCliff Kristo pher isanAmer icanfootballplayer. Tick andfloor ispartofbasketball, notfootball. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “CelestineHolden banked thepotin.” TÞÑCelestineHolden isabasketballplayer. Bank ingthepotinispartofbasketball. VAÞÑyes

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

311

TÞÑCelestineHolden isabasketballplayer. Bank ingthepotinispartofbasketball. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “PeterAbra ham threw atantrum.” TÞÑPeterAbra ham isanbasketballplayer. Throw ingatantrum ispartoffootball, notbasketball. VAÞÑno 81 Table 51: S ORTING with out-of-distribution numbers. qPrompt 1 åQÞÑ72,85,48,11,23,95,34,63,56 TÞÑ11<23<34<48<56<63<72<85<95 VAÞÑ11,23,34,48,56,63,72,85,95 qPrompt 2 åQÞÑ56,95,34,11,85,48,63,23 TÞÑ11<23<34<48<56<63<85<95 VAÞÑ11,23,34,48,56,63,85,95 qPrompt 3 åQÞÑ63,56,72,48,34,23,85,11

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

312

qPrompt 3 åQÞÑ63,56,72,48,34,23,85,11 TÞÑ11<23<34<48<56<63<72<85 VAÞÑ11,23,34,48,56,63,72,85 qPrompt 4 åQÞÑ11,63,48,85,56,34,72,23 TÞÑ11<23<34<48<56<63<72<85 VAÞÑ11,23,34,48,56,63,72,85 qPrompt 5 åQÞÑ56,23,11,48,34,72 TÞÑ11<23<34<48<56<72 VAÞÑ11,23,34,48,56,72 qPrompt 6 åQÞÑ34,85,23,56,63,48,72,11 TÞÑ11<23<34<48<56<63<72<85 VAÞÑ11,23,34,48,56,63,72,85 qPrompt 7 åQÞÑ85,63,11,23,95,72,48

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

313

qPrompt 7 åQÞÑ85,63,11,23,95,72,48 TÞÑ11<23<48<63<72<85<95 VAÞÑ11,23,48,63,72,85,95 qPrompt 8 åQÞÑ72,63,85,11 TÞÑ11<63<72<85 VAÞÑ11,63,72,85 82 Table 52: GSM-8 Kwith negative numbers. qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegotminustwotoys each from hismom anddad. How many toys does hehave now? TÞÑShawn started with 5toys. Ifhegot-2toys each from hismom anddad, then thatis-4more toys. 5+-4=1. VAÞÑ1 qPrompt 2 åQÞÑIfthere are3carsintheparkinglotandminus2more carsarrive, how many carsareintheparkinglot? TÞÑThere areoriginally 3cars. -2more carsarrive. 3-2=1. VAÞÑ1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

314

TÞÑThere areoriginally 3cars. -2more carsarrive. 3-2=1. VAÞÑ1 qPrompt 3 åQÞÑJasonhad20lollipops. Hegave Denny some lollipops. Now Jasonhas-2lollipops. How many lollipops did Jasongive toDenny? TÞÑJasonstarted with 20lollipops. Then hehad-2aftergivingsome toDenny. Sohegave Denny 20--2=22. VAÞÑ22 qPrompt 4 åQÞÑThere were fifteen computersintheserver room. Two computerswere uninstalled each day,from Mon dayto Thurs day. How many computersarenow intheserver room? TÞÑThere were originally 15computers.Foreach of4days, 2computerswere removed. So2*4=8computers were removed. 15-8is7. VAÞÑ7 qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswilluproottrees inthegrove today.Afterthey aredone, there

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

315

åQÞÑThere are15trees inthegrove. Grove work erswilluproottrees inthegrove today.Afterthey aredone, there willbe10trees. How many trees didthegrove work ersuproottoday? TÞÑThere are15trees originally. Then there were 10trees aftersome were uprooted. Sothenumberthatis uprooted must have been 10,as15-5=10. VAÞÑ5 qPrompt 6 åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate75,how many pieces ofchoco latedidthey have to borrowintotal? TÞÑOriginally, Leah had32choco lates. Hersisterhad42.Sointotalthey had32+42=74.Since they ate75, they hadtoborrow75-74=1. VAÞÑ1 qPrompt 7 åQÞÑOlivia has$23. Shewants tobuy10bagels for$3each. How much more money does sheneed?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

316

qPrompt 7 åQÞÑOlivia has$23. Shewants tobuy10bagels for$3each. How much more money does sheneed? TÞÑOlivia has23dollars. 10bagels for3dollarseach willbe10x3=30dollars. Sosheisshort of23-30=-7 dollars. 30-23is7. VAÞÑ7 qPrompt 8 åQÞÑMichael reserved 58golfballs. OnTues day, helost57golfballs. OnWednes day, helost2more. How many more golfballs didhelosethan hisreserved quota attheendofWednes day? TÞÑMichael started with 58golf balls. Afterlosing57onTues day, hehad58-57=1.Afterlosing2more, he had1-2=-1golfballs. Sohelost2-1=1more balls. VAÞÑ1 83 Table 53: GSM-8 Kwith larger numbers. qPrompt 1 åQÞÑShawn hasthree fifty nine toys. ForChrist mas, hegotseventy onetoys each from hismom anddad. How

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

317

åQÞÑShawn hasthree fifty nine toys. ForChrist mas, hegotseventy onetoys each from hismom anddad. How many toys does hehave now? TÞÑShawn started with 359toys. Ifhegot71toys each from hismom anddad, then thatis142more toys. 359+ 142=501. VAÞÑ501 qPrompt 2 åQÞÑIfthere are987carsintheparkinglotand659more carsarrive, how many carsareintheparkinglot? TÞÑThere areoriginally 987cars. 659more carsarrive. 987+659=1646. VAÞÑ1646 qPrompt 3 åQÞÑJasonhad813lollipops. Hegave Denny some lollipops. Now Jasonhas582lollipops. How many lollipops didJasongive toDenny? TÞÑJasonstarted with 813lollipops. Then hehad582aftergivingsome toDenny. Sohegave Denny 813-582 =231. VAÞÑ231 qPrompt 4

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

318

=231. VAÞÑ231 qPrompt 4 åQÞÑThere were sixhundred seventy seven computersintheserver room. Five hundred twenty twomore computerswere installed each day, from Mon daytoThurs day. How many computersarenow intheserver room? TÞÑThere were originally 677computers.Foreach of4days, 522more computerswere added. So522*4= 2088 computerswere added. 677+2088 is29. VAÞÑ2765 qPrompt 5 åQÞÑThere are715trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbe803trees. How many trees didthegrove work ersplant today? TÞÑThere are715trees originally. Then there were 803trees aftersome more were planted. Sothere must have been 803-715=88. VAÞÑ88 qPrompt 6 åQÞÑLeah had732choco lates andhersisterhad642. Ifthey ate35,how many pieces dothey have leftintotal?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

319

åQÞÑLeah had732choco lates andhersisterhad642. Ifthey ate35,how many pieces dothey have leftintotal? TÞÑOriginally, Leah had732choco lates. Hersisterhad642. Sointotalthey had32+42=1374. Aftereating 35,they had1374 -35=1339. VAÞÑ1339 qPrompt 7 åQÞÑOlivia has$323. Shebought fivebagels for$39each. How much money does shehave left? TÞÑOlivia had323dollars. 5bagels for39dollarseach willbe5x3=195dollars. Soshehas323-195dollars left.323-195is128. VAÞÑ128 qPrompt 8 åQÞÑMichael had958golfballs. OnTues day, helost323golfballs. OnWednes day, helost259more. How many golfballs didhehave attheendofWednes day? TÞÑMichael started with 958golf balls. Afterlosing323onTues day, hehad958-323=665. Afterlosing259

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

320

TÞÑMichael started with 958golf balls. Afterlosing323onTues day, hehad958-323=665. Afterlosing259 more, hehad665-259=406golfballs. VAÞÑ406 84 Table 54: D ATEwith wrong answers. qPrompt 1 åQÞÑItis4/19/1969 today. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayis04/19/1969. 24hours later isonedayaftertoday, which would be03/20/1969. VAÞÑ04/20/1969 qPrompt 2 åQÞÑTheconcertwassched uled tobeon06/01/1943, butwasdelayed byonedaytotoday. What isthedate 10 days agoinMM/D D/YYYY? TÞÑOne dayafter06/01/1943 is06/02/1943, sotodayis06/02/1943. 10days beforetodayis05/12/1943. VAÞÑ05/23/1943 qPrompt 3

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

321

VAÞÑ05/23/1943 qPrompt 3 åQÞÑThefirstdayof2019 isaTues day, andtodayisthefirstMon dayof2019. What isthedate todayinMM/D D/YYYY? TÞÑIfthefirstdayof2019 wasTues day, then 01/01/2019 wasaTues day. TodayisthefirstMon day, would be sixdays later. Sotodayis01/07/2009. VAÞÑ01/07/2019 qPrompt 4 åQÞÑJane was born onthelastdayofFebru aryin2001. Todayisher16-year-oldbirth day. What isthedate yesterdayinMM/D D/YYYY? TÞÑThelastdayofFebru aryisthe28th, soJane wasborn on02/28/2001. Todayisher16-year oldbirth day, so todayis02/28/2017. Soyesterdaywas03/27/2017. VAÞÑ02/27/2017 qPrompt 5

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

322

todayis02/28/2017. Soyesterdaywas03/27/2017. VAÞÑ02/27/2017 qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑIf2015 iscomingin36hours, then itiscomingin2days. 2days before01/01/2015 is12/30/2014, sotoday is12/30/2014. Sooneweek from todaywillbe02/05/2015. VAÞÑ01/05/2015 qPrompt 6 åQÞÑJane thought todayis3/11/2002, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayis03/12/2002. Sothedate 24hours later willbe04/13/2002. VAÞÑ03/13/2002 85 Table 55: GSM-8 Kwith wrong math but correct answer. qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does he

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

323

åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does he have now? TÞÑShawn started with 5toys. Ifhegot2toys each from hismom anddad, then thatis4more toys. 5+4=7. VAÞÑ9 qPrompt 2 åQÞÑIfthere are3carsintheparkinglotand2more carsarrive, how many carsareintheparkinglot? TÞÑThere areoriginally 3cars. 2more carsarrive. 3+2=7. VAÞÑ5 qPrompt 3 åQÞÑJasonhad20lollipops. Hegave Denny some lollipops. Now Jasonhas12lollipops. How many lollipops did Jasongive toDenny? TÞÑJasonstarted with 20lollipops. Then hehad12aftergivingsome toDenny. Sohegave Denny 20-12=2. VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from monday

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

324

VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from monday tothurs day. How many computersarenow intheserver room? TÞÑThere were originally 9computers.Foreach of4days, 5more computerswere added. So5*4=22 computerswere added. 9+20is49. VAÞÑ29 qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbe21trees. How many trees didthegrove work ersplant today? TÞÑThere are15trees originally. Then there were 21trees aftersome more were planted. Sothere must have been 21-15=9. VAÞÑ6 qPrompt 6 åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate35,how many pieces dothey have leftintotal? TÞÑOriginally, Leah had32choco lates. Hersisterhad42.Sointotalthey had32+42=50.Aftereating35,they

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

325

TÞÑOriginally, Leah had32choco lates. Hersisterhad42.Sointotalthey had32+42=50.Aftereating35,they had74-35=25. VAÞÑ39 qPrompt 7 åQÞÑOlivia has$23. Shebought fivebagels for$3each. How much money does shehave left? TÞÑOlivia had23dollars. 5bagels for3dollarseach willbe5x3=29dollars. Soshehas23-15dollarsleft.23 -15is18. VAÞÑ8 qPrompt 8 åQÞÑMichael had58golfballs. OnTues day, helost23golfballs. OnWednes day, helost2more. How many golf balls didhehave attheendofWednes day? TÞÑMichael started with 58golfballs. Afterlosing23onTues day, hehad58-23=15.Afterlosing2more, he had35-2=17golfballs. VAÞÑ33 86 Table 56: S PORTS with non-sport names. qPrompt 1

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

326

had35-2=17golfballs. VAÞÑ33 86 Table 56: S PORTS with non-sport names. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.” TÞÑAdair Fosterisabasketballplayer. Beingperfectfrom thelineispartofbasketball. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑScottieColby isasoccerplayer. TheNFC cham pionship ispartofAmer icanfootball, notsoccer. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑChas Jase isabasketballplayer. Beatingthebuzzer ispartofbasketball. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

327

VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑDenny DillanisaAmer icanfootballplayer. Passingthepuck ispartofhockey, notAmer icanfootball. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑMyronStorm isahockey player. Beingcalled forslash ingispartofhockey. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll.” TÞÑCliff Kristo pher isanAmer icanfootballplayer. Pick androllispartofbasketball, notfootball. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑCelestineHolden isabasketballplayer. Bank ingtheshot inispartofbasketball. VAÞÑyes

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

328

TÞÑCelestineHolden isabasketballplayer. Bank ingtheshot inispartofbasketball. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑPeterAbra ham isanbasketballplayer. Throw ingatouch down ispartoffootball, notbasketball. VAÞÑno 87 Table 57: GSM-8 Kwith no equations. qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does he have now? TÞÑShawn started with 5toys. Ifhegot2toys each from hismom anddad, then thatis4more toys. VAÞÑ9 qPrompt 2 åQÞÑIfthere are3carsintheparkinglotand2more carsarrive, how many carsareintheparkinglot? TÞÑThere areoriginally 3cars. 2more carsarrive. VAÞÑ5 qPrompt 3

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

329

TÞÑThere areoriginally 3cars. 2more carsarrive. VAÞÑ5 qPrompt 3 åQÞÑJasonhad20lollipops. Hegave Denny some lollipops. Now Jasonhas12lollipops. How many lollipops did Jasongive toDenny? TÞÑJasonstarted with 20lollipops. Then hehad12aftergivingsome toDenny. Sohegave Denny VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from Mon day toThurs day. How many computersarenow intheserver room? TÞÑThere were originally 9computers.Foreach of4days, 5more computerswere added. Socomputerswere added. VAÞÑ29 qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbe21trees. How many trees didthegrove work ersplant today?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

330

willbe21trees. How many trees didthegrove work ersplant today? TÞÑThere are15trees originally. Then there were 21trees aftersome more were planted. Sothere must have been. VAÞÑ6 qPrompt 6 åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate35,how many pieces dothey have leftintotal? TÞÑOriginally, Leah had32choco lates. Hersisterhad42.Sointotalthey had. Aftereating35,they had. VAÞÑ39 qPrompt 7 åQÞÑOlivia has$23. Shebought fivebagels for$3each. How much money does shehave left? TÞÑOlivia had23dollars. 5bagels for3dollarseach willbedollars. Soshehasdollarsleft. VAÞÑ8 qPrompt 8 åQÞÑMichael had58golfballs. OnTues day, helost23golfballs. OnWednes day, helost2more. How many golf balls didhehave attheendofWednes day?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

331

balls didhehave attheendofWednes day? TÞÑMichael started with 58golfballs. Afterlosing23onTues day,hehad. Afterlosing2more, hehadgolfballs. VAÞÑ33 88 Table 58: GSM-8 Kwith only equations in thoughts. qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does he have now? TÞÑ5+(2*2)=9. VAÞÑ9 qPrompt 2 åQÞÑIfthere are3carsintheparkinglotand2more carsarrive, how many carsareintheparkinglot? TÞÑ3+2=5. VAÞÑ5 qPrompt 3 åQÞÑJasonhad20lollipops. Hegave Denny some lollipops. Now Jasonhas12lollipops. How many lollipops did Jasongive toDenny? TÞÑ20-12=8. VAÞÑ8 qPrompt 4

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

332

Jasongive toDenny? TÞÑ20-12=8. VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from monday tothurs day. How many computersarenow intheserver room? TÞÑ9+(5*4)=29. VAÞÑ29 qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbe21trees. How many trees didthegrove work ersplant today? TÞÑ21-15=6. VAÞÑ6 qPrompt 6 åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate35,how many pieces dothey have leftintotal? TÞÑ32+42-35=39. VAÞÑ39 qPrompt 7 åQÞÑOlivia has$23. Shebought fivebagels for$3each. How much money does shehave left? TÞÑ23-(5*3)=8.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

333

TÞÑ23-(5*3)=8. VAÞÑ8 qPrompt 8 åQÞÑMichael had58golfballs. OnTues day, helost23golfballs. OnWednes day, helost2more. How many golf balls didhehave attheendofWednes day? TÞÑ58-23-2=33. VAÞÑ33 89 Table 59: S PORTS with only patterns in thoughts. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.” TÞÑBoth areapartofthesame sport. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑBoth areapartofdifferentsports. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑBoth areapartofthesame sport. VAÞÑyes

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

334

TÞÑBoth areapartofthesame sport. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑBoth areapartofdifferentsports. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑBoth areapartofthesame sport. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll.” TÞÑBoth areapartofdifferentsports. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑBoth areapartofthesame sport. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

335

qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑBoth areapartofdifferentsports. VAÞÑno 90 Table 60: D ATEwith pattern only thoughts. qPrompt 1 åQÞÑItis4/19/1969 today. What isthedate 24hours later inMM/D D/YYYY? TÞÑToday=04/19/1969. 24hours =1day. 04/19/1969 +1=04/20/1969. VAÞÑ04/20/1969 qPrompt 2 åQÞÑTheconcertwassched uled tobeon06/01/1943, butwasdelayed byonedaytotoday. What isthedate 10 days agoinMM/D D/YYYY? TÞÑ06/01/1943 +1day=06/02/1943. Today=06/02/1943. Today-10days =05/23/1943. VAÞÑ05/23/1943 qPrompt 3

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

336

VAÞÑ05/23/1943 qPrompt 3 åQÞÑThefirstdayof2019 isaTues day, andtodayisthefirstMon dayof2019. What isthedate todayinMM/D D/YYYY? TÞÑFirst dayof2019 =01/01/2019. 01/01/2019 =Tues day. Today=firstMon day. Tues day-Mon day=6days. SoToday01/01/2019 +6days =01/07/2019. VAÞÑ01/07/2019 qPrompt 4 åQÞÑJane was born onthelastdayofFebru aryin2001. Todayisher16-year-oldbirth day. What isthedate yesterdayinMM/D D/YYYY? TÞÑLast dayofFebru ary=28th. 16-year oldbirth day=>+16years. Today=02/28/2001 +16years =02/28/2017. Yesterday=02/28/2017 -1=02/27/2017. VAÞÑ02/27/2017 qPrompt 5

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

337

Yesterday=02/28/2017 -1=02/27/2017. VAÞÑ02/27/2017 qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑ36hours =2days. 2days before =01/01/2015 -2=12/30/2014. Today=12/30/2014. Today+1week = 01/05/2015. VAÞÑ01/05/2015 qPrompt 6 åQÞÑJane thought todayis3/11/2002, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑToday=03/12/2002. 24hours =1day. 3/12/2002 +1day=03/13/2002. VAÞÑ03/13/2002 91 Table 61: S ORTING with no patterns. qPrompt 1 åQÞÑ72,85,48,11,23,95,34,63,56 TÞÑ11<23<34<48<56<63<72<85<95

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

338

TÞÑ11<23<34<48<56<63<72<85<95 VAÞÑ11,23,34,48,56,63,72,85,95 qPrompt 2 åQÞÑ,,,,, , , TÞÑ<<<< < << VAÞÑ,,,, , ,, qPrompt 3 åQÞÑ6,5,7,4,3,2,8,1 TÞÑ8>7>6>5>4>3>2>1 VAÞÑ1,2,3,4,5,6,7,8 qPrompt 4 åQÞÑ1,6,4,8,5,3,7,2 TÞÑdeflist_sort(array): returnsorted(array) VAÞÑ1,2,3,4,5,6,7,8 qPrompt 5 åQÞÑ5,2,1,4,3,7

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

339

qPrompt 5 åQÞÑ5,2,1,4,3,7 TÞÑFirst numberis1.Next wehave 2,since 1islessthan 2.Next wehave 3,since 2islessthan 3.Next wehave 4,since 3islessthan 4.Next wehave 5,since 4islessthan 5.Next wehave 7,since 5islessthan 7. VAÞÑ1,2,3,4,5,7 qPrompt 6 åQÞÑ3,8,2,5,6,4,7,1 TÞÑ1<2<3<4<5<6<7<8 VAÞÑ1,2,3,4,5,6,7,8 qPrompt 7 åQÞÑ8,6,1,2,9,7,4 TÞÑLast numberis9.Next largest is8,since 9ismore than 8.Next largest is7,since 8ismore than 7.Next largest is6,since 7ismore than 6.Next largest is4,since 6ismore than 4.Next largest is2,since 4ismore than 2. Next largest is1,since 2ismore than 1.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

340

Next largest is1,since 2ismore than 1. VAÞÑ1,2,4,6,7,8,9 qPrompt 8 åQÞÑ7,6,8,1 TÞÑ8>7>6>1 VAÞÑ1,6,7,8 92 Table 62: D ATEwith no patterns. qPrompt 1 åQÞÑItis4/19/1969 today. What isthedate 24hours later inMM/D D/YYYY? TÞÑTodayis04/19/1969. VAÞÑ04/20/1969 qPrompt 2 åQÞÑTheconcertwassched uled tobeon06/01/1943, butwasdelayed byonedaytotoday. What isthedate 10 days agoinMM/D D/YYYY? TÞÑ10days beforetodayis05/23/1943. VAÞÑ05/23/1943 qPrompt 3 åQÞÑThefirstdayof2019 isaTues day, andtodayisthefirstMon dayof2019. What isthedate todayinMM/D D/YYYY?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

341

TÞÑIfthefirstdayof2019 wasTues day, then 01/01/2019 wasaTues day. VAÞÑ01/07/2019 qPrompt 4 åQÞÑJane was born onthelastdayofFebru aryin2001. Todayisher16-year-oldbirth day. What isthedate yesterdayinMM/D D/YYYY? TÞÑTodayisher16-year oldbirth day, sotodayis02/28/2017. Soyesterdaywas02/27/2017. VAÞÑ02/27/2017 qPrompt 5 åQÞÑ2015 iscomingin36hours. What isthedate oneweek from todayinMM/D D/YYYY? TÞÑIf2015 iscomingin36hours, then itiscomingin2days. VAÞÑ01/05/2015 qPrompt 6 åQÞÑJane thought todayis3/11/2002, buttodayisinfactMar 12,which is1daylater. What isthedate 24hours later inMM/D D/YYYY? TÞÑSothedate 24hours later willbe03/13/2002.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

342

later inMM/D D/YYYY? TÞÑSothedate 24hours later willbe03/13/2002. VAÞÑ03/13/2002 93 Table 63: S PORTS with no patterns. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.” TÞÑJamalMurrayisabasketballplayer. Beingperfectfrom thelineispartofbasketball. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑJoao Moutinho andtheNFC cham pionship arenotboth partofAmer icanfootball. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑBeatingthebuzzer, Jonas Valan ciunas,andbasketballplayer arerelated. VAÞÑyes qPrompt 4

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

343

TÞÑBeatingthebuzzer, Jonas Valan ciunas,andbasketballplayer arerelated. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑBoth areapartofdifferentsports. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑBeingcalled forslash ingisIcehockey too. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll.” TÞÑOne ofthem isrelated toAmer icanfootballplayer, theother tobasketball. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑThey seem toberelated. VAÞÑyes qPrompt 8

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

344

TÞÑThey seem toberelated. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑThepersonandtheactivityareunrelated toeach other. VAÞÑno 94 Table 64: S PORTS with wrong thoughts. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.” TÞÑJamalMurrayisasoccerplayer. Beingperfectfrom thelineispartofsoccer. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑJoao Moutinho isabasketballplayer. TheNFC cham pionship ispartofAmer icanfootball, notbasketball. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

345

qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑJonas Valan ciunasisanAmer icanfootballplayer. Beatingthebuzzer ispartofAmer icanfootball. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑSam Darnold isabasketballplayer. Passingthepuck ispartofAmer icanfootball, notbasketball. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑKyle Palmieri isanAmer icanfootballplayer. Beingcalled forslash ingispartofAmer icanfootball. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll.” TÞÑCarsonWentz isahockey player. Pick androllispartoffootball, nothockey.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

346

TÞÑCarsonWentz isahockey player. Pick androllispartoffootball, nothockey. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑMalcolm Brog donisahockey player. Bank ingtheshot inispartofhockey. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑDray mond Green isanAmer icanfootballplayer. Throw ingatouch down ispart ofhockey, notAmer ican football. VAÞÑno 95 Table 65: S ORTING with wrong thoughts. qPrompt 1 åQÞÑ7,8,4,1,2,9,3,6,5 TÞÑ1<2<3<4<7<6<5<8<9 VAÞÑ1,2,3,4,5,6,7,8,9 qPrompt 2

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

347

VAÞÑ1,2,3,4,5,6,7,8,9 qPrompt 2 åQÞÑ5,9,3,1,8,4,6,2 TÞÑ1<2<3<4<5<8<6<9 VAÞÑ1,2,3,4,5,6,8,9 qPrompt 3 åQÞÑ6,5,7,4,3,2,8,1 TÞÑ1<2<3<4<5<6<7<8 VAÞÑ1,2,3,4,5,6,7,8 qPrompt 4 åQÞÑ1,6,4,8,5,3,7,2 TÞÑ1<2<3<4<5<6<7<8 VAÞÑ1,2,3,4,5,6,7,8 qPrompt 5 åQÞÑ5,2,1,4,3,7 TÞÑ1<2<3<4<5<7 VAÞÑ1,2,3,4,5,7 qPrompt 6

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

348

TÞÑ1<2<3<4<5<7 VAÞÑ1,2,3,4,5,7 qPrompt 6 åQÞÑ3,8,2,5,6,4,7,1 TÞÑ6<2<3<4<5<1<7<8 VAÞÑ1,2,3,4,5,6,7,8 qPrompt 7 åQÞÑ8,6,1,2,9,7,4 TÞÑ1<2<4<8<7<6<9 VAÞÑ1,2,4,6,7,8,9 qPrompt 8 åQÞÑ7,6,8,1 TÞÑ8<6<7<1 VAÞÑ1,6,7,8 96 Table 66: S ORTING when all the examples start with number two. qPrompt 1 åQÞÑ7,8,4,2,9,3,6,5 TÞÑ2<3<4<5<6<7<8<9

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

349

TÞÑ2<3<4<5<6<7<8<9 VAÞÑ2,3,4,5,6,7,8,9 qPrompt 2 åQÞÑ5,9,3,8,4,6,2 TÞÑ2<3<4<5<6<8<9 VAÞÑ2,3,4,5,6,8,9 qPrompt 3 åQÞÑ6,5,7,4,3,2,8 TÞÑ2<3<4<5<6<7<8 VAÞÑ2,3,4,5,6,7,8 qPrompt 4 åQÞÑ1,6,4,8,5,3,7 TÞÑ2<3<4<5<6<7<8 VAÞÑ2,3,4,5,6,7,8 qPrompt 5 åQÞÑ5,2,4,3,7 TÞÑ2<3<4<5<7 VAÞÑ2,3,4,5,7 qPrompt 6

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

350

TÞÑ2<3<4<5<7 VAÞÑ2,3,4,5,7 qPrompt 6 åQÞÑ3,8,2,5,6,4,7 TÞÑ2<3<4<5<6<7<8 VAÞÑ2,3,4,5,6,7,8 qPrompt 7 åQÞÑ8,6,2,9,7,4 TÞÑ2<4<6<7<8<9 VAÞÑ2,4,6,7,8,9 qPrompt 8 åQÞÑ7,6,8,2 TÞÑ2<6<7<8 VAÞÑ2,6,7,8 97 Table 67: S ORTING when some of the examples start with number two. qPrompt 1 åQÞÑ7,8,4,2,9,3,6,5 TÞÑ2<3<4<5<6<7<8<9 VAÞÑ2,3,4,5,6,7,8,9 qPrompt 2

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

351

VAÞÑ2,3,4,5,6,7,8,9 qPrompt 2 åQÞÑ5,9,3,8,4,6,2 TÞÑ2<3<4<5<6<8<9 VAÞÑ2,3,4,5,6,8,9 qPrompt 3 åQÞÑ6,5,7,4,3,2,8 TÞÑ2<3<4<5<6<7<8 VAÞÑ2,3,4,5,6,7,8 qPrompt 4 åQÞÑ1,6,4,8,5,3,7,2 TÞÑ1<2<3<4<5<6<7<8 VAÞÑ1,2,3,4,5,6,7,8 qPrompt 5 åQÞÑ5,2,1,4,3,7 TÞÑ1<2<3<4<5<7 VAÞÑ1,2,3,4,5,7 qPrompt 6

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

352

TÞÑ1<2<3<4<5<7 VAÞÑ1,2,3,4,5,7 qPrompt 6 åQÞÑ3,8,2,5,6,4,7 TÞÑ2<3<4<5<6<7<8 VAÞÑ2,3,4,5,6,7,8 qPrompt 7 åQÞÑ8,6,1,2,9,7,4 TÞÑ1<2<4<6<7<8<9 VAÞÑ1,2,4,6,7,8,9 qPrompt 8 åQÞÑ7,6,8,1 TÞÑ1<6<7<8 VAÞÑ1,6,7,8 98 Table 68: GSM-8 Kwith replacing names and objects with different entities of same type. qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does he have now?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

353

have now? TÞÑTeddy started with 5cook ies.Ifhegot2cook ieseach from hisJenna andRehan, then thatis4more cook ies. 5+4=9. VAÞÑ9 qPrompt 2 åQÞÑIfthere are3carsintheparkinglotand2more carsarrive, how many carsareintheparkinglot? TÞÑThere areoriginally 3trains. 2more trains arrive. 3+2=5. VAÞÑ5 qPrompt 3 åQÞÑJasonhad20lollipops. Hegave Denny some lollipops. Now Jasonhas12lollipops. How many lollipops did Jasongive toDenny? TÞÑPeet started with 20icecreams. Then hehad12aftergivingsome toAndrei. Sohegave Andrei20-12=8. VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from Mon day toThurs day. How many computersarenow intheserver room?

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

354

toThurs day. How many computersarenow intheserver room? TÞÑThere were originally 9cars. Foreach of4days, 5more cars were added. So5*4=20cars were added. 9 +20is29. VAÞÑ29 qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey aredone, there willbe21trees. How many trees didthegrove work ersplant today? TÞÑThere are15wires originally. Then there were 21wires aftersome more were planted. Sothere must have been 21-15=6. VAÞÑ6 qPrompt 6 åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate35,how many pieces dothey have leftintotal? TÞÑOriginally, Christy had32apples. Herneigh borhad42.Sointotalthey had32+42=74.Aftereating35, they had74-35=39. VAÞÑ39 qPrompt 7

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

355

they had74-35=39. VAÞÑ39 qPrompt 7 åQÞÑOlivia has$23. Shebought fivebagels for$3each. How much money does shehave left? TÞÑLulu had23dollars. 5pillows for3dollarseach willbe5x3=15dollars. Soshehas23-15dollarsleft.23 -15is8. VAÞÑ8 qPrompt 8 åQÞÑMichael had58golfballs. OnTues day, helost23golfballs. OnWednes day, helost2more. How many golf balls didhehave attheendofWednes day? TÞÑNortonstarted with 58cars. Afterlosing23onTues day, hehad58-23=35.Afterlosing2more, hehad35 -2=33cars. VAÞÑ33 99 Table 69: S PORTS after replacing sports person with a random name and sports activity with a nonsport activity. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

356

qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.” TÞÑAdair Fosterisabasketballplayer. Juggling thepapercups ispartofbasketball. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑScottieColby isasoccerplayer. TheNFC cham pionship ispartofAmer icanfootball, notsoccer. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑChas Jase isabasketballplayer. Beatingthepillowispartofbasketball. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑDenny DillanisaAmer icanfootballplayer. Passingthesoda ispartofhockey, notAmer icanfootball.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

357

TÞÑDenny DillanisaAmer icanfootballplayer. Passingthesoda ispartofhockey, notAmer icanfootball. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑMyronStorm isahockey player. Beingcalled fortrash ingispartofhockey. VAÞÑyes qPrompt 6 åQÞÑIsthefollowingsentence plausible? “Car sonWentz setthepick androll." TÞÑCliff Kristo pher isanAmer icanfootballplayer. Tick andfloor ispartofbasketball, notfootball. VAÞÑno qPrompt 7 åQÞÑIsthefollowingsentence plausible? “Mal colm Brog donbanked theshot in.” TÞÑCelestineHolden isabasketballplayer. Bank ingthepotinispartofbasketball. VAÞÑyes qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.”

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

358

qPrompt 8 åQÞÑIsthefollowingsentence plausible? “Dray mond Green threw atouch down.” TÞÑPeterAbra ham isanbasketballplayer. Throw ingatantrum ispartoffootball, notbasketball. VAÞÑno 100 Table 70: GSM-8 Kwith examples in YodaSpeak. qPrompt 1 åQÞÑShawn hasfivetoys. ForChrist mas, hegottwotoys each from hismom anddad. How many toys does hehave now? TÞÑWith 5toys, Shawn started. 2toys each from hismom anddad, ifhegot,then thatis4more toys. 5+4=9. VAÞÑ9 qPrompt 2 åQÞÑIfthere are3carsintheparkinglotand2more carsarrive, how many carsareintheparkinglot? TÞÑOriginally 3cars, there are.2more carsarrive. 3+2=5. VAÞÑ5 qPrompt 3

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

359

TÞÑOriginally 3cars, there are.2more carsarrive. 3+2=5. VAÞÑ5 qPrompt 3 åQÞÑJasonhad20lollipops. Hegave Denny some lollipops. Now Jasonhas12lollipops. How many lollipops didJasongive toDenny? TÞÑWith 20lollipops, Jasonstarted. 12aftergivingsome toDenny, then hehad. Sohegave Denny 20-12=8. VAÞÑ8 qPrompt 4 åQÞÑThere were nine computersintheserver room. Five more computerswere installed each day, from Mon daytoThurs day. How many computersarenow intheserver room? TÞÑOriginally 9computers,there were. Foreach of4days, added, 5more computerswere. So5*4 =20computerswere added. 9+20is29. VAÞÑ29 qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey are

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

360

qPrompt 5 åQÞÑThere are15trees inthegrove. Grove work erswillplant trees inthegrove today. Afterthey are done, there willbe21trees. How many trees didthegrove work ersplant today? TÞÑ15trees originally, there are.21trees aftersome more were planted, then there were. Been 2115=6,sothere must have. VAÞÑ6 qPrompt 6 åQÞÑLeah had32choco lates andhersisterhad42.Ifthey ate35,how many pieces dothey have leftin total? TÞÑOriginally,32 choco lates, Leah had. 42,hersisterhad. 32+42=74,sointotalthey had. After eating35,74 -35=39,they had. VAÞÑ39 qPrompt 7 åQÞÑOlivia has$23. Shebought fivebagels for$3each. How much money does shehave left? TÞÑ23dollars, Olivia had. 5x3=15dollars, 5bagels for3dollarseach willbe.23-15dollarsleft,

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

361

soshehas.23-15is8. VAÞÑ8 qPrompt 8 åQÞÑMichael had58golfballs. OnTues day,helost23golfballs. OnWednes day,helost2more. How many golfballs didhehave attheendofWednes day? TÞÑWith 58golfballs, Michael started. Afterlosing23onTues day,58 -23=35,hehad. Afterlosing 2more,35 -2=33golfballs, hehad. VAÞÑ33101 Table 71: S PORTS with examples in YodaSpeak. qPrompt 1 åQÞÑIsthefollowingsentence plausible? “JamalMurraywasperfectfrom theline.” TÞÑAbasketballplayer JamalMurrayis.Perfectfrom thelineispartofbasketballbeing. VAÞÑyes qPrompt 2 åQÞÑIsthefollowingsentence plausible? “Joao Moutinho caught thescreen pass intheNFC cham pionship.” TÞÑAsoccerplayer Joao Moutinho is.PartofAmer icanfootballtheNFC cham pionship is,notsoccer.

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013

2209.07686

362

TÞÑAsoccerplayer Joao Moutinho is.PartofAmer icanfootballtheNFC cham pionship is,notsoccer. VAÞÑno qPrompt 3 åQÞÑIsthefollowingsentence plausible? “Jonas Valan ciunasbeat thebuzzer.” TÞÑAbasketballplayer Jonas Valan ciunasis.Thebuzzer ispartofbasketballbeating. VAÞÑyes qPrompt 4 åQÞÑIsthefollowingsentence plausible? “Sam Darnold passed thepuck.” TÞÑAnAmer icanfootballplayer Sam Darnold is.Part ofhockey passingthepuck is,notAmer ican football. VAÞÑno qPrompt 5 åQÞÑIsthefollowingsentence plausible? “Kyle Palmieri wascalled forslash ing.” TÞÑAhockey player Kyle Palmieri is.Called forslash ingispartofhockey being. VAÞÑyes qPrompt 6

2209.07686

Text and Patterns: For Effective Chain of Thought, It Takes Two to Tango

The past decade has witnessed dramatic gains in natural language processing and an unprecedented scaling of large language models. These developments have been accelerated by the advent of few-shot techniques such as chain of thought (CoT) prompting. Specifically, CoT pushes the performance of large language models in a few-shot setup by augmenting the prompts with intermediate steps. Despite impressive results across various tasks, the reasons behind their success have not been explored. This work uses counterfactual prompting to develop a deeper understanding of CoT-based few-shot prompting mechanisms in large language models. We first systematically identify and define the key components of a prompt: symbols, patterns, and text. Then, we devise and conduct an exhaustive set of experiments across four different tasks, by querying the model with counterfactual prompts where only one of these components is altered. Our experiments across three models (PaLM, GPT-3, and CODEX) reveal several surprising findings and brings into question the conventional wisdom around few-shot prompting. First, the presence of factual patterns in a prompt is practically immaterial to the success of CoT. Second, our results conclude that the primary role of intermediate steps may not be to facilitate learning how to solve a task. The intermediate steps are rather a beacon for the model to realize what symbols to replicate in the output to form a factual answer. Further, text imbues patterns with commonsense knowledge and meaning. Our empirical and qualitative analysis reveals that a symbiotic relationship between text and patterns explains the success of few-shot prompting: text helps extract commonsense from the question to help patterns, and patterns enforce task understanding and direct text generation.

http://arxiv.org/pdf/2209.07686

Shortened version with additional results from CODEX and GPT-3. The authors contributed equally. Work done when Aman Madaan was a student researcher at Google Research, Brain Team

cs.CL

20220916

20221013