id
stringlengths 12
15
| title
stringlengths 8
162
| content
stringlengths 1
17.6k
| prechunk_id
stringlengths 0
15
| postchunk_id
stringlengths 0
15
| arxiv_id
stringlengths 10
10
| references
sequencelengths 1
1
|
|---|---|---|---|---|---|---|
2310.02255#153 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: From the given figure: Circle + Square = 5 =â Circle = 5 â Square . . . (i) Triangle + Triangle = 8 =â Triangle = 4 . . . (ii) Circle + Triangle = 7 Using (ii), Circle + 4 = 7 =â Circle = 3 . . . (iii) Using (i) and (iii), 5 â Square = 3 =â Square = 2. Correct output: 2 Figure 66: The question assesses abstract thinking through a puzzle test problem. Only GPT-4V grasps the abstract concepts and employs step-by-step reasoning to deduce the final answer. 63 | 2310.02255#152 | 2310.02255#154 | 2310.02255 | [
"2302.13971"
] |
2310.02255#154 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Published as a conference paper at ICLR 2024 ID: 778 Question: Can you find the missing shape in this picture puzzle? Choices: (A) A (B) B (C) C (D) D Ground truth: (A) A LLaMA-Adapter-V2: The missing shape in this picture puzzle is a hexagon. LLaVA: The missing shape in the picture puzzle is a square. Multimodal Bard: The missing shape in the picture puzzle is (D). The shapes in the puzzle are all regular polygons, and the number of sides increases by 1 as you move from left to right. The first shape is a pentagon, the second shape is a hexagon, the third shape is an octagon, and the fourth shape is a decagon. The fifth shape should be a dodecagon. The answer choice A is a pentagon, which is already in the puzzle. The answer choice B is a heptagon, which has 7 sides. The answer choice C is a nonagon, which has 9 sides. The answer choice D is a dodecagon, which has 12 sides. Therefore, the missing shape is the dodecagon. | 2310.02255#153 | 2310.02255#155 | 2310.02255 | [
"2302.13971"
] |
2310.02255#155 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: The pattern seems to involve alternating filled and unfilled shapes. Following the pattern: - The first shape is a filled hexagon. - The second shape is an unfilled hexagon. Thus, the missing shape should be a filled trapezoid. The correct option is: (D) D. Figure 67: All baselines struggle to discern the shapes presented in the problem and the options, and are unable to identify the correct pattern. 64 Published as a conference paper at ICLR 2024 ID: 750 Question: Which number is missing? Ground truth: 18 Solution: 5 Ã 2 = 10, 6 Ã 2 = 10, 9 Ã 2 = 18 12 10 2 (5) (2) (6) LLaMA-Adapter-V2: The missing number is 4. LLaVA: The number missing in the image is 3. Multimodal Bard: The missing number in the puzzle is 7. The numbers in the puzzle are arranged in a circle, and each number is surrounded by the sum of the two numbers next to it. For example, the number 12 is surrounded by 9 and 10, because 9 + 10 = 19. The missing number must be surrounded by 2 and 5, and the only number that satisfies this condition is 7. Therefore, the missing number is 7. | 2310.02255#154 | 2310.02255#156 | 2310.02255 | [
"2302.13971"
] |
2310.02255#156 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: From the given image, the number in the center, 2, is the result of the difference between the number directly opposite it in each pair: 12 â 10 = 2 9 â 5 = 4 6â ? = 2 Given that the difference between 6 and the missing number is 2, the missing number is: 6 â 2 = 4. So, the missing number is 4. Figure 68: All baselines fail to deduce the pattern inherent in the logical reasoning problem. 65 | 2310.02255#155 | 2310.02255#157 | 2310.02255 | [
"2302.13971"
] |
2310.02255#157 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Published as a conference paper at ICLR 2024 H.3.5 NUMERIC COMMONSENSE REASONING Problems involving numeric commonsense reasoning on MATHVISTA require commonsense knowl- edge about daily objects and celebrities to answer visual questions. However, these problems present significant challenges to existing foundation models, including GPT-4V, as depicted in Figure 1. For instance, Multimodal Bard struggles to understand the optical illusion in an image (Figure 69) and to infer the age gap between two celebrities from another image (Figure 70). Figure 71 poses a question about the maximum volume a beaker can measure. However, GPT-4V lacks commonsense knowledge regarding the use of a beaker, resulting in an incorrect prediction. ID: 693 Question: | 2310.02255#156 | 2310.02255#158 | 2310.02255 | [
"2302.13971"
] |
2310.02255#158 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Is the orange larger than the car? Choices: (A) Yes (B) No Ground truth: (B) No LLaMA-Adapter-V2: No, the orange is not larger than the car. Correct output: (B) No LLaVA: The orange is larger than the car. Multimodal Bard: The answer is (B), No. The orange is not larger than the car. The car is clearly much larger than the orange. Even if it is a large orange, it is still much smaller than the car. The orange is probably about the size of a basketball, while the car is at least the size of a small SUV. | 2310.02255#157 | 2310.02255#159 | 2310.02255 | [
"2302.13971"
] |
2310.02255#159 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | I canâ t see the orange in the image you sent, but I can still answer the question based on the information you provided. The question is asking if the orange is larger than the car in general, not in the specific image. So, even if the orange in the image is larger than the car in the image, the answer would still be no, because oranges are not generally larger than cars. Correct output: (B) No GPT-4V: The orange is not larger than the car; itâ s just closer to the camera, creating an optical illusion. | 2310.02255#158 | 2310.02255#160 | 2310.02255 | [
"2302.13971"
] |
2310.02255#160 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Correct output: (B) No Figure 69: The question pertains to everyday common sense and the identification of ob- jectsâ specifically an orange and a carâ in the image. GPT-4V produces the correct answer by discerning an optical illusion present in the image. 66 Published as a conference paper at ICLR 2024 ID: 29 Question: What is the age gap between these two peo- ple in image? (unit: years) Ground truth: 4 Ay f) }| a fy ya ~ WG £ LLaMA-Adapter-V2: | 2310.02255#159 | 2310.02255#161 | 2310.02255 | [
"2302.13971"
] |
2310.02255#161 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | The age gap between these two people in the image is 10 years. LLaVA: The age gap between the two people in the image is 20 years. Multimodal Bard: Based on the image you sent, the age gap between the king and queen is 3 years. The king is Richard III of England, who was born in 1452. The queen is Anne Neville, who was born in 1456. So, Richard was 3 years older than Anne when they married in 1472. The stained glass window is likely depicting the couple after their marriage, so the age gap between them would still be 3 years. | 2310.02255#160 | 2310.02255#162 | 2310.02255 | [
"2302.13971"
] |
2310.02255#162 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: The two figures depicted in the stained glass are King Richard III and Anne Neville. Richard III was born on October 2, 1452, and Anne Neville was born on June 11, 1456. Calculating the age gap: 1456 â 1452 = 4 So, the age gap between Richard III and Anne Neville is 4 years. Correct output: 4 Figure 70: In this instance, only GPT-4V correctly recognizes the celebrities along with their profiles to calculate the age gap. 67 | 2310.02255#161 | 2310.02255#163 | 2310.02255 | [
"2302.13971"
] |
2310.02255#163 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Published as a conference paper at ICLR 2024 ID: 19 Question: What glass measures? Ground truth: 400 is the highest amount this LLaMA-Adapter-V2: The highest amount this glass measures is 400 ml. Correct output: 400 LLaVA: The highest amount this glass measures is 3000 ml. Multimodal Bard: The highest amount the beaker in the image measures is 600 ml. The markings on the side of the beaker are in milliliters (ml). The largest marking is 600 ml, which is the maximum amount of liquid that the beaker can hold. The other markings on the beaker are 300 ml, 200 ml, 100 ml, and -400 ml. The negative markings indicate the amount of liquid that is below the zero mark. | 2310.02255#162 | 2310.02255#164 | 2310.02255 | [
"2302.13971"
] |
2310.02255#164 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: The highest amount this glass measures is 600 ml. Figure 71: Most AI models without tool augmentation fall short in this example. A model not only needs to recognize the text and numbers accurately but also to possess numerical common sense to determine the highest amount that the beaker can measure. 68 Published as a conference paper at ICLR 2024 # H.3.6 SCIENTIFIC REASONING Scientific reasoning represents a distinct mathematical reasoning ability within our MATHVISTA. To tackle problems in this area, a model must not only accurately interpret domain-specific information from figures, but also possess the necessary in-domain knowledge to reason rigorously on scientific topics. Figure 1 shows that GPT-4V substantially outperforms the other foundation models. This superiority is further illustrated by the examples in Figures 72 and 73. However, the failure of GPT-4V, as shown in Figure 74, indicates that there is considerable room for improvement. | 2310.02255#163 | 2310.02255#165 | 2310.02255 | [
"2302.13971"
] |
2310.02255#165 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | ID: 807 Question: Which of the following leaf shapes would have the least amount of wind resistance and water loss? Choices: (A) Truncate (B) Acuminate (C) Rounded (D) Sagittate Ground truth: (B) Acuminate acuminate WV oblique . acute sagittate rounded hastate cordate Up truncate ©Em.Arstrong 2002 LLaMA-Adapter-V2: The correct answer is (D) Sagittate.. LLaVA: The leaf shape with the least amount of wind resistance and water loss would be the truncate leaf. This is because truncate leaves have a more streamlined shape, which reduces air resistance and minimizes water loss due to evaporation. | 2310.02255#164 | 2310.02255#166 | 2310.02255 | [
"2302.13971"
] |
2310.02255#166 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Multimodal Bard: The answer is (C). Rounded leaves have the least amount of wind resistance and water loss. This is because they have a smooth, curved surface that does not catch the wind as much as other leaf shapes. Rounded leaves also have a smaller surface area than other leaf shapes, which means that they lose less water through evaporation. The other leaf shapes in the choices are not as aerodynamic as rounded leaves. Truncate leaves have a flat end, which can catch the wind. Acuminate leaves have a long, pointed tip, which can also catch the wind. Sagittate leaves have a long, arrow-shaped base, which can also catch the wind. | 2310.02255#165 | 2310.02255#167 | 2310.02255 | [
"2302.13971"
] |
2310.02255#167 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: The leaf shape that would have the least amount of wind resistance and water loss would be the one that is most slender and pointed. This is because a slender and pointed shape would allow wind to flow around it with the least resistance, and a narrow surface area would lead to less water loss through transpiration. Among the options, the â acuminateâ leaf shape is the most slender and pointed. Correct output: (B) Acuminate Figure 72: The question requires grasping biology concepts depicted in the image. LLaMA- Adapter-V2, LLaVA, and Multimodal Bard demonstrate a deficiency in domain-specific knowledge and struggle to interpret the illustration. | 2310.02255#166 | 2310.02255#168 | 2310.02255 | [
"2302.13971"
] |
2310.02255#168 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 69 Published as a conference paper at ICLR 2024 Table 5. View interpolation results (PSNR) for NVDIFFREC reconstructions of the NeRF synthetic dataset, using either DMTEt or FLexiCuses for the topology step. The image metric scores are arithmetic means over all test images. We also include Chamfer distances (CD) computed on visible trian- gles (the set of triangles visible in at least one test view) using 2.5 M point. Lower scores indicate better geometric fidelity. PSNR (dB) 1 Chair Drums Ficus Hotdog Lego Mats Mic Ship DMTer 31.8 246 30.9 33.2 29.0 27.0 30.7 26.0 FLEXICUBES 31.8 24.7 30.9 33.4 28.8 26.7 30.8 25.9 CD (10-2) | Chair Drums Ficus Hotdog Lego Mats Mic Ship DMTeT 451 3.98 0.30 267 2.41 0.41 1.20 55.8 FLEXICUBES 0.45 2.27 0.37 144 1.60 0.53 1.51 10.5 PSNR (dB) 1 Chair Drums Ficus Hotdog Lego Mats Mic Ship DMTer 31.8 246 30.9 33.2 29.0 27.0 30.7 26.0 FLEXICUBES 31.8 24.7 30.9 33.4 28.8 26.7 30.8 25.9 CD (10-2) | Chair Drums Ficus Hotdog Lego Mats Mic Ship DMTeT 451 3.98 0.30 267 2.41 0.41 1.20 55.8 FLEXICUBES 0.45 2.27 0.37 144 1.60 0.53 1.51 10.5 # ID: 278 Question: Among the following objects, which one has the best PSNR score? # Choices: (A) Lego (B)Mats (C) Mic (D)Ship | 2310.02255#167 | 2310.02255#169 | 2310.02255 | [
"2302.13971"
] |
2310.02255#169 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | # Ground truth: (C) Mic LLaMA-Adapter-V2: The correct option letter for the best PSNR score is â C,â which is the PSNR score for the mic. LLaVA: The correct answer is (D) Ship. # Multimodal Bard: The answer is (D) Ship. The PSNR score for the ship is 33.4 dB, which is higher than the PSNR scores for the other objects. The PSNR score measures the amount of distortion between the original image and the reconstructed image. A higher PSNR score indicates less distortion and a better quality reconstruction. The other objects have the following PSNR scores: * Lego: 29.0 dB * Mats: 27.0 dB * Mic: 30.7 dB Therefore, the ship has the best PSNR score among the given objects. | 2310.02255#168 | 2310.02255#170 | 2310.02255 | [
"2302.13971"
] |
2310.02255#170 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: Based on the table for PSNR (dB) scores: Lego: DMTet: 29.0 FLeXiCubes: 28.8 Mats: DMTet: 27.0 FLeXiCubes: 26.7 Mic: DMTet: 30.7 FLeXiCubes: 30.8 Ship: DMTet: 26.0 FLeXiCubes: 25.9 Among the choices, â Micâ has the highest PSNR scores with both DMTet and FLeXiCubes, so the answer is (C) Mic. Correct output: (C) Mic Figure 73: The problem necessitates a unified comprehension of the table and its corresponding caption within academic domains. Exclusively, GPT-4V accurately recognizes the OCR text from related columns and deduces the correct answer. | 2310.02255#169 | 2310.02255#171 | 2310.02255 | [
"2302.13971"
] |
2310.02255#171 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 70 Published as a conference paper at ICLR 2024 ID: 725 Question: In Figure, suppose that Barbaraâ s velocity rela- tive to Alex is a constant vBA = 52 km/h and car P is moving in the negative direction of the x axis. If Alex measures a constant vP A = â 78 km/h for car P , what velocity vP B will Barbara measure? Ground truth: -130 Frame B moves past frame A while both observe P. Frame B | 2310.02255#170 | 2310.02255#172 | 2310.02255 | [
"2302.13971"
] |
2310.02255#172 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | LLaMA-Adapter-V2: Barbara will measure a velocity of vP B = â 52 km/h relative to car P . LLaVA: If Alex measures a constant vP A = â 78 km/h for car P , Barbara will measure a velocity of vP B = â 26 km/h relative to car P . Multimodal Bard: The velocity of car P relative to Barbara is given by: vP B = vP A â vBA = â 78 km/h â 52 km/h = â 130 km/h Therefore, Barbara will measure a velocity of -130 km/h for car P . Here is the explanation: * The velocity of car P relative to Alex, vP A, is measured by Alex, who is stationary in frame A. * The velocity of Barbara relative to Alex, vBA, is measured by Alex, who is also stationary in frame A. * The velocity of car P relative to Barbara, vP B, is measured by Barbara, who is moving in frame B with a velocity of vBA. According to the relative velocity addition rule, the velocity of car P relative to Barbara is equal to the velocity of car P relative to Alex minus the velocity of Barbara relative to Alex. In this case, the velocity of car P relative to Alex is -78 km/h and the velocity of Barbara relative to Alex is 52 km/h. Therefore, the velocity of car P relative to Barbara is -130 km/h. Correct output: -130 GPT-4V: Given: vBA = 52 km/h (velocity of Barbara relative to Alex) vP A = -78 km/h (velocity of car P as measured by Alex) We need to find vP B (velocity of car P as measured by Barbara). Using the Galilean transformation for velocities in one dimension: vP B = vP A + vBA vP B = â 78 km/h + 52 km/h vP B = â 26 km/h Figure 74: This college-level question evaluates the capacity to understand scientific figure, com- mand domain-specific knowledge in science, and execute rigorous reasoning. While GPT-4V mis- applies its domain-specific knowledge resulting in an incorrect response, Multimodal Bard delivers the correct answer supported by a logical explanation. | 2310.02255#171 | 2310.02255#173 | 2310.02255 | [
"2302.13971"
] |
2310.02255#173 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 71 Published as a conference paper at ICLR 2024 H.3.7 STATISTICAL REASONING In MATHVISTA, problems encompass a variety of charts, plots, and graphs designed to assess the statistical reasoning capabilities of foundation models. As demonstrated in Figure 1, GPT-4V shows strong statistical reasoning ability. For instance, GPT-4V produces accurate answers for the format- rich table in Figure 75 and the data analysis table in Figure 76. | 2310.02255#172 | 2310.02255#174 | 2310.02255 | [
"2302.13971"
] |
2310.02255#174 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | ID: 649 Question: How many values are below 30 in Mainly are incidents of individual misconduct? Ground truth: 4 Most Americans say reports of sexual misconduct reflect societal problems = % who say recent allegations of sexual harassment and assault... Mainly are Mai incidents of individual â misconduct Men Women College grad+ Some college HS or less Rep/Lean Rep Dem/Lean Dem LLaMA-Adapter-V2: There are 2 values below 30 in Mainly are incidents of individual misconduct. LLaVA: There are two values below 30 in Mainly are incidents of individual misconduct. Multimodal Bard: There is only 1 value below 30 in the â Mainly are incidents of individual misconductâ column. This value is 22, which is the percentage of women who believe that recent allegations of sexual harassment and assault are mainly incidents of individual misconduct. The other values in this column are 34, 31, 33, and 25. All of these values are greater than 30. So the answer to your question is 1. | 2310.02255#173 | 2310.02255#175 | 2310.02255 | [
"2302.13971"
] |
2310.02255#175 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: Based on the â Mainly are incidents of individual misconductâ column, the following values are below 30: - 28 (Total) - 22 (Women) - 19 (College grad+) - 25 (Dem/Lean Dem) Thatâ s a total of 4 values. So, the answer is 4. Correct output: 4 Figure 75: The question with a richly formatted statistical figure presents challenges for prominent multimodal models such as LLaMA-Adapter-V2, LLaVA, and Multimodal Bard. However, GPT-4V successfully extracts essential information from the figure, producing the correct answer. | 2310.02255#174 | 2310.02255#176 | 2310.02255 | [
"2302.13971"
] |
2310.02255#176 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 72 Published as a conference paper at ICLR 2024 ID: 389 Question: Look at the table. Then answer the question. At a price of $320, is there a shortage or a surplus? Choices: (A) shortage (B) surplus Ground truth: (A) shortage LLaMA-Adapter-V2: The table shows a surplus of 1000 units at a price of $320. LLaVA: At a price of $320, there is a surplus of 12,000 units. | 2310.02255#175 | 2310.02255#177 | 2310.02255 | [
"2302.13971"
] |
2310.02255#177 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Multimodal Bard: The correct answer is (B), surplus. At a price of $320, the quantity supplied is 12,500 units, while the quantity demanded is 20,000 units. This means that there are 7,500 more units supplied than demanded. So the answer is surplus. Here is a table summarizing the supply and demand at each price: Price Quantity Supplied Quantity Demanded $270 $320 $370 $420 $470 10,600 12,500 14,400 19,600 18,200 20,200 20,000 19,800 19,400 16,300 Surplus/Shortage Shortage of 9,600 Surplus of 7,500 Surplus of 5,400 No surplus or shortage Shortage of 1,900 | 2310.02255#176 | 2310.02255#178 | 2310.02255 | [
"2302.13971"
] |
2310.02255#178 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: At a price of $320: Quantity demanded = 20,000 Quantity supplied = 12,500 Since the quantity demanded (20,000) is greater than the quantity supplied (12,500), there is a shortage. Correct output: (A) shortage Figure 76: LLaMA-Adapter-V2, LLaVA, and Multimodal Bard struggle to discern the structured information within the table, leading to incorrect answers. 73 Published as a conference paper at ICLR 2024 H.4 ABILITIES ACROSS VISUAL CONTEXTS This section compares the reasoning abilities of GPT-4V with other large multimodal models (LLMs) on MATHVISTA, considering various types of visual contexts. Models used for compar- ison include LLaMA-Adapter-V2-7B (simplified as LLaMA-Adapter-V2), LLaVA-LLaMA-2-13B (simplified as LLaVA), and Multimodal Bard. H.4.1 ABSTRACT SCENE Based on Figure 1, current foundation models lag behind human performance in mathematical rea- soning in abstract scenes by a substantial margin. Consider the problems in Figures 77 and 78 that are derived from math word problems found in elementary school curricula. Despite their advanced capabilities, foundation models such as Multimodal Bard and GPT-4V fail to produce the correct responses. ID: 742 Question: | 2310.02255#177 | 2310.02255#179 | 2310.02255 | [
"2302.13971"
] |
2310.02255#179 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Move the ruler to measure the length of the twig to the nearest inch. The twig is about ( ) inches long. Ground truth: 3 LLaMA-Adapter-V2: The twig is about 2 inches long. LLaVA: The twig is about 2 inches long. Multimodal Bard: According to the image, the 0 inch mark of the ruler is aligned with the base of the twig. The tip of the twig is just past the 5 inch mark. Therefore, the twig is about 5 inches long. | 2310.02255#178 | 2310.02255#180 | 2310.02255 | [
"2302.13971"
] |
2310.02255#180 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: The twig is about 4 inches long. â â â PaaS Sr Seaeeee Figure 77: Large multimodal models like GPT-4V struggle to answer this elementary-level question with the abstract scene as the context. 74 Published as a conference paper at ICLR 2024 ID: 7 Question: What time is shown? Answer by typing a time word, not a num- ber. It is ( ) after eight. Choices: (A) half (B) quarter (C) oâ clock (D) quarter to (E) quarter past Ground truth: (B) quarter | 2310.02255#179 | 2310.02255#181 | 2310.02255 | [
"2302.13971"
] |
2310.02255#181 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | LLaMA-Adapter-V2: The time shown is â half past eight.â LLaVA: The correct answer is (C) oâ clock. Multimodal Bard: The image you sent shows a clock with the hands pointing to 8:15. So the time is quarter past eight. (E) The hour hand is between 8 and 9, which means it is past 8 oâ clock. The minute hand is pointing to the 3, which is 15 minutes past. Therefore, the time is quarter past eight. | 2310.02255#180 | 2310.02255#182 | 2310.02255 | [
"2302.13971"
] |
2310.02255#182 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: The minute hand is pointing to the 9, which indicates that it is 45 minutes past the hour. The hour hand is pointing a little past the 8, which means itâ s after 8 oâ clock. 45 minutes past the hour is also described as â quarter toâ the next hour. So, the time shown is â quarter to nine.â The correct answer is (D) quarter to. Figure 78: Large multimodal models face challenges in answering this elementary-level question which entails interpreting the time from an abstract clock diagram. 75 | 2310.02255#181 | 2310.02255#183 | 2310.02255 | [
"2302.13971"
] |
2310.02255#183 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Published as a conference paper at ICLR 2024 H.4.2 BAR CHART As shown in Figure 1, foundation models, including GPT-4V, significantly underperform humans in mathematical reasoning when bar charts serve as the visual context. Neither Multimodal Bard nor GPT-4V can solve the problems depicted in Figures 79 and 80, which do not need complex understanding and reasoning. ID: 497 Question: What is the value of the largest bar? Ground truth: 9 LLaMA-Adapter-V2: The value of the largest bar is 10. LLaVA: The value of the largest bar is 90. Multimodal Bard: The largest bar in the image is the one for the title â ninetyâ . It has a value of 10. GPT-4V: The value of the largest bar is 10. Figure 79: Large Multimodal Models such Multimodal Bard and GPT-4 fail to extract the correct value for the tallest bar in this basic bar chart. ID: 629 Question: What percentage of people prefer the object hut? Ground truth: 20 LLaMA-Adapter-V2: 10% of people prefer the object hut.. LLaVA: Based on the image, 30% of people prefer the object hut.. Most preferred objects Ed 0 6 eo 100 Percent of People | 2310.02255#182 | 2310.02255#184 | 2310.02255 | [
"2302.13971"
] |
2310.02255#184 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Multimodal Bard: Based on the image, 100% of people prefer the object hut. The line graph shows that the percentage of people who prefer the hut is 100%, which is the maximum value on the graph. GPT-4V: The percentage of people who prefer the object â hutâ is approximately 40%. Figure 80: To address this question, the model simply needs to identify the â hutâ and determine its value on the x-axis. However, current foundation models such as GPT-4V struggle to produce the correct answer. | 2310.02255#183 | 2310.02255#185 | 2310.02255 | [
"2302.13971"
] |
2310.02255#185 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 76 Published as a conference paper at ICLR 2024 H.4.3 FUNCTION PLOT GPT-4V outperforms other baselines on problems related to function plots and even exceeds human performance. Figures 81 and 82 show questions with digital and hand-drawn function plots, respec- tively. In both cases, GPT-4V accurately identifies their functions and infers the correct properties. ID: 767 Question: What is the maximum value of y? Ground truth: 5 2 2 x" +Y" = 25 (3, -4) LLaMA-Adapter-V2: The maximum value of y is 3. LLaVA: The maximum value of y is 3. Multimodal Bard: The maximum value of y is 3. The given equation is y2 + x2 = 25 point (3, â 4) lies on the circle, so the maximum value of y is 3. 2 . This is the equation of a circle with center (0, 0) and radius 5â 2 . The GPT-4V: | 2310.02255#184 | 2310.02255#186 | 2310.02255 | [
"2302.13971"
] |
2310.02255#186 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | The equation given is for a circle: x2 + y2 = 25. The radius of this circle is the square root of 25, which is 5. The maximum value of y corresponds to the highest point of the circle, which is equal to the radius. So, the maximum value of y is 5. Correct output: 5 Figure 81: The figure illustrates a function plot accompanied by its equation. GPT-4V correctly identifies the equation, leading to the correct answer. | 2310.02255#185 | 2310.02255#187 | 2310.02255 | [
"2302.13971"
] |
2310.02255#187 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 77 Published as a conference paper at ICLR 2024 ID: 39 Question: Is this function odd or even? Choices: (A) odd (B) even Ground truth: (A) odd LLaMA-Adapter-V2: The function is odd. Correct output: (A) odd LLaVA: The function in the image is odd. Correct output: (A) odd Multimodal Bard: The function in the image is odd. An even function is a function where f (â x) = f (x). An odd function is a function where f (â x) = â f (x). In the case of the function y = a2x â a1, if we substitute x = â x, we get: y = a2(â x) â a1 = â a2x + a1 This is not equal to f (x), so the function is odd. The correct choice is (A). Correct output: (A) odd GPT-4V: The given function is y = ax where a > 1. To determine if the function is odd or even, we can use the following definitions: A function f (x) is even if f (â x) = f (x) for all x in the domain of f . - A function f (x) is odd if f (â x) = â f (x) for all x in the domain of f . For the given function y = ax, when we plug in â x for x: # y = aâ x This is not equal to â ax and itâ s not equal to ax either. Thus, the function is neither odd nor even. However, since the given choices only include â oddâ and â evenâ | 2310.02255#186 | 2310.02255#188 | 2310.02255 | [
"2302.13971"
] |
2310.02255#188 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | , the best answer is neither. But since I must choose one of the provided options (A) odd. Correct output: (A) odd Figure 82: The figure showcases a hand-drawn function plot. GPT-4V identifies the functionâ s expression and accurately determines its property. 78 Published as a conference paper at ICLR 2024 H.4.4 GEOMETRY DIAGRAM Geometry diagrams are a distinct type of visual context in MATHVISTA. To answer questions involv- ing these diagrams, a model must comprehend the fine-grained details, including symbols, variables, and relations from the figures. Additionally, it should apply associated theorems before executing calculations to produce final responses. GPT-4V surpasses other models and even humans due to its superior capabilities in geometry recognition and reasoning. In the examples shown in Figures 83 and 84, GPT-4V delivers the correct results through the application of relevant theorems and subsequent calculations. | 2310.02255#187 | 2310.02255#189 | 2310.02255 | [
"2302.13971"
] |
2310.02255#189 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | ID: 478 Question: As shown in the figure, in the diamond ABCD, two diagonal lines AC = 12.0, BD = 16.0, then the edge length of this diamond is (). Choices: (A) 10 (B) 8 (C) 6 (D) 5 Ground truth: 10 P LLaMA-Adapter-V2: The edge length of this diamond is 8. LLaVA: Based on the given information, the edge length of the diamond ABCD is 8.0. # Multimodal Bard: The answer is (A). 10. In a rhombus, the diagonals bisect each other at right angles. So, we can draw two right triangles within the rhombus, each with a leg of length 6 and a hypotenuse of length 10. | 2310.02255#188 | 2310.02255#190 | 2310.02255 | [
"2302.13971"
] |
2310.02255#190 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | â â By the Pythagorean Theorem, the length of each side of the rhombus is 62 + 62 = 72 = 10. So the answer is (A). Correct output: 10 GPT-4V: Given that AC = 12.0 and BD = 16.0 in the rhombus (diamond) ABCD. Since all sides of a rhombus are congruent, the diagonals of a rhombus are perpendicular bisectors of each other. This means that each half of the diagonal BD is 8.0 and each half of diagonal AC is 6.0. Using the Pythagorean theorem for one of the right triangles (formed by the diagonals): Side2 = (Half of BD)2 + (Half of AC)2 Side2 = 82 + 62 Side2 = 64 + 36 Side2 = 100 Side = 10 # Correct output: 10 Figure 83: For this geometry problem, GPT-4V correctly applies the appropriate theorem and arrives at the right answer. In contrast, Multimodal Bard commits calculation errors during its reasoning process. 79 Published as a conference paper at ICLR 2024 ID: 410 Question: Find the perimeter of â ³DEF, if â ³DEF â ¼ â ³CBF, perimeter of â ³CBF = 27, DF = 6, and F C = 8. Choices: (A) 20.25 (B) 21 (C) 27 (D) 36 Ground truth: (A) 20.25 c B_ LLaMA-Adapter-V2: The perimeter of â ³DEF is 20.25, as shown in the image. Correct output: 20.25 LLaVA: The perimeter of â ³DEF can be calculated using the similarity theorem, which states that if two triangles are similar, their corresponding sides are proportional. In this case, we know that â ³DEF â ¼ â ³CBF and the perimeter of â ³CBF is 27. Since the ratio of the corresponding sides of similar triangles is the same, we can use the perimeter of â ³CBF to find the perimeter of â ³DEF . The perimeter of â ³DEF is proportional to the perimeter of â ³CBF , so we can multiply the perimeter of â ³CBF by the ratio of the corresponding sides. | 2310.02255#189 | 2310.02255#191 | 2310.02255 | [
"2302.13971"
] |
2310.02255#191 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | The ratio of the corresponding sides of â ³DEF and â ³CBF is 6:8, which means that the perimeter of â ³DEF is 6 Ã 8 = 48. Therefore, the correct answer is (D) 48. # Multimodal Bard: Since â ³DEF â ¼ â ³CBF , we have DE # CB = DF F C . Since DE = 6 and F C = 8, we have DE # CB = 6 8 = 3 4 . The perimeter of â ³CBF is 27, so CB + BF = 27. Substituting 3 Solving for BF , we get BF = 12. 4 for CB BF , we get 3 4 BF + BF = 27. The perimeter of â ³DEF is DE + EF + DF = 6 + EF + 6 = EF + 12. Substituting 3 get EF + 12 = 4 Therefore, the perimeter of â ³DEF is EF + DE + DF = 9 + 6 + 6 = 21 . GPT-4V: Given â ³DEF â ¼ â ³CBF and DF = 6 and F C = 8, we can find the scale factor of the two similar triangles. # Scale factor = DF # F C = 6 8 = 3 4 . Given the perimeter of â ³CBF = 27, we can find the perimeter of â ³DEF by multiplying by the scale factor. Perimeter of â ³DEF = 27 Ã 3 Correct output: 20.25 4 = 20.25. Figure 84: For this geometry problem, GPT-4V correctly applies the appropriate theorem and arrives at the right answer. In contrast, Multimodal Bard commits calculation errors during its reasoning process. | 2310.02255#190 | 2310.02255#192 | 2310.02255 | [
"2302.13971"
] |
2310.02255#192 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 80 Published as a conference paper at ICLR 2024 H.4.5 LINE PLOT As evidenced by Figure 1, current models such as GPT-4V do not perform as well as humans in mathematical reasoning involving line plots. We speculate that the low performance is mainly due to the difficulty in detecting OCR text in the figures and accurately grounding the values, as illustrated by the examples in Figures 85 and 86. ID: 981 Question: What 2002, 2003 and 2004? is the sum of Ground truth: 70.4 LLaMA-Adapter-V2: The sum of 2002, 2003 and 2004 is 6.6. LLaVA: The sum of 2002, 2003, and 2004 is 6007. Multimodal Bard: | 2310.02255#191 | 2310.02255#193 | 2310.02255 | [
"2302.13971"
] |
2310.02255#193 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | I see the image. The sum of 2002, 2003 and 2004 is 1.40. The year 2002 has a share of 67.4%, 2003 has a share of 64.3%, and 2004 has a share of 62%. When we add these three values, we get 1.937%. Rounding to one decimal place, the sum is 1.40. GPT-4V: The sum for the years 2002, 2003, and 2004 from the graph are: 2002: 19.4% 2003: 23.2% 2004: 25.1% Sum = 19.4 + 23.2 + 25.1 = 67.7% Figure 85: The figure displays a low-resolution line plot with numerous points. Current foundation models, including GPT-4V, appear to struggle with accurately interpreting the plot. 81 Published as a conference paper at ICLR 2024 ID: 209 Question: What est value in black line chart? is the high- Ground truth: 28.3 LLaMA-Adapter-V2: The highest value in the black line chart is 1.4. LLaVA: The highest value in the black line chart is 1.4. Multimodal Bard: The highest value in the black line chart is 1.4. 30.1 po a9 72329 2 7259 26.8 269269 26. 2 25.9259 955 25:6: 25.3 25.3 25.3 25.3 25 25.1.25.125.1 GPT-4V: The highest value in the black line chart, which represents â Women,â is 30.1. Figure 86: The figure displays two low-resolution line plots comprising numerous points. Current foundation models, including GPT-4V, seem to struggle with interpreting the plots accurately. | 2310.02255#192 | 2310.02255#194 | 2310.02255 | [
"2302.13971"
] |
2310.02255#194 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 82 Published as a conference paper at ICLR 2024 H.4.6 NATURAL IMAGE MATHVISTA includes questions that require numeric and spatial reasoning based on text and objects in natural images. If models have limited abilities to recognize text (OCR), as shown in Figure 87, or to identify visual objects, as in Figure 88, they are unlikely to generate correct answers to visual questions. ID: 980 Question: What is the highest number shown? Ground truth: 12 LLaMA-Adapter-V2: The highest number shown is 12. Correct output: 12 LLaVA: The highest number shown on the clock tower is 3. | 2310.02255#193 | 2310.02255#195 | 2310.02255 | [
"2302.13971"
] |
2310.02255#195 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Multimodal Bard: I see the image. The highest number shown is 3. GPT-4V: The highest number shown on the clock is 12. Correct output: 12 Figure 87: LLaVA and Multimodal Bard are unable to identify the highest number displayed on the clock panel in the image. 83 Published as a conference paper at ICLR 2024 ID: 884 Question: What is the blue kite in the lower right corner shaped like? Choices: (A) ferret (B) cat (C) cloud (D) octopus Ground truth: (D) octopus LLaMA-Adapter-V2: The blue kite in the lower right corner is shaped like a cloud. | 2310.02255#194 | 2310.02255#196 | 2310.02255 | [
"2302.13971"
] |
2310.02255#196 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | LLaVA: The blue kite in the lower right corner is shaped like a cloud. Multimodal Bard: Sorry, I canâ t help with images of people yet. GPT-4V: The blue kite in the lower right corner is shaped like (D) octopus. Correct output: (D) octopus Figure 88: The question demands spatial reasoning ability from the model to pinpoint the object within the image. In this context, only GPT-4V delivers a correct prediction. 84 Published as a conference paper at ICLR 2024 H.4.7 PUZZLE TEST Math reasoning with puzzle text figures is challenging for current AI foundation models because interpreting these figures requires discerning underlying patterns from sets of shapes, as illustrated in Figure 89, and numbers, as in Figure 90. | 2310.02255#195 | 2310.02255#197 | 2310.02255 | [
"2302.13971"
] |
2310.02255#197 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | There is plenty of room for improvement. ID: 123 Question: The diagram represents successive rotations, starting from the top down. Which shape comes next? Choices: (A) A (B) B (C) C (D) D (E) E Ground truth: (D) D LLaMA-Adapter-V2: The diagram represents successive rotations, starting from the top down. The next shape that comes next is letter D. Correct output: (D) D LLaVA: The correct answer is (E) E. # Multimodal Bard: The answer is (C). The diagram shows a 3D cube being rotated 90 degrees on each axis. Starting from the top, the first shape is a square. The second shape is a cube with a square missing from the top face. The third shape is a cube with a square missing from the front face. The fourth shape is a cube with a square missing from the left face. To find the next shape, we need to rotate the cube 90 degrees on the z-axis. This will bring the square that is missing from the left face to the front face. So the next shape is the one shown in option (C). GPT-4V: The diagram shows a pattern of rotating the shape in a specific sequence. To determine the next shape: Starting with the first shape, when we rotate it 90 degrees counterclockwise, we get the second shape. Rotating the second shape 90 degrees counterclockwise gives the third shape. Thus, rotating the third shape 90 degrees counterclockwise will give us the next shape. From the choices provided, the shape that matches this rotation is (E) E. Figure 89: This puzzle test figure demands intricate skills in abstract thinking and spatial reasoning, which prove challenging for current foundation models. Although GPT-4V correctly discerns the transformation pattern of the shapes, it errs in spatially distinguishing between shapes (D) and (E). | 2310.02255#196 | 2310.02255#198 | 2310.02255 | [
"2302.13971"
] |
2310.02255#198 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 85 Published as a conference paper at ICLR 2024 ID: 941 Question: Which is the next number in the series? Ground truth: 2 Solution: (3968 + 1) = 632 (63 + 1) = 82 (8 + 1) = 32 (3 + 1) = 22 LLaMA-Adapter-V2: The next number in the series is 3. LLaVA: The next number in the series is 4. | 2310.02255#197 | 2310.02255#199 | 2310.02255 | [
"2302.13971"
] |
2310.02255#199 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Multimodal Bard: The next number in the series is 9. The series is based on the square root of the number plus 1. So, the first number is 3968, and its square root is 63. Then, 63 + 1 = 64, and its square root is 8. Continuing the pattern, 8 + 1 = 9, and its square root is 3. Therefore, the next number in the series is 9. | 2310.02255#198 | 2310.02255#200 | 2310.02255 | [
"2302.13971"
] |
2310.02255#200 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: The pattern in the ladder involves dividing the number by the step number to get the next number. Step 1: 3968 ÷ 1 = 3968 Step 2: 3968 ÷ 2 = 1984 but we have 63, which doesnâ t follow this pattern. Step 3: 63 ÷ 3 = 21 but we have 8. Step 4: 8 ÷ 4 = 2 but we have 3. Step 5: 3 ÷ 5 = 0.6 The given sequence doesnâ t seem to have a consistent pattern based on the values presented. However, if we assume the ladder sequence is based on dividing by a consistent factor (rather than step number), we can attempt to find a common factor: 3968 ÷ 63â 63 63 ÷ 8= 7.875 which isnâ t a round number. 8 ÷ 3= 2.6666 again not a round number. Without a clear and consistent pattern, itâ s challenging to determine the next number in the sequence. Figure 90: The question demands advanced abstract thinking to discern the consistent pattern within a sequence of four numbers. Despite their capabilities, current foundation models including Multi- modal Bard and GPT-4V show limited abilities of logical reasoning. 86 Published as a conference paper at ICLR 2024 H.4.8 SCATTER PLOT A scatter plot is a graphical representation of data points on a two-dimensional axis, where each point represents a value from the dataset. MATHVISTA includes the reasoning task that requires comprehending scatter plots taken from daily life and academic papers, as shown in Figures 91 and 92. Although GPT-4V outperforms other LMMs, such as Multimodal Bard, and even humans in overall accuracy (Figure 1), it often fails in the cases where fine-grained understanding is required, as in Figure 92. ID: 105 Question: Does Dark Violet have the minimum area under the curve? Choices: (A) yes (B) no Ground truth: (A) yes Dark Violet Hot Pink Medium Mint Dark Cyan Yellow Green a & xaxis_label LLaMA-Adapter-V2: No, Dark Violet does not have the minimum area under the curve. LLaVA: | 2310.02255#199 | 2310.02255#201 | 2310.02255 | [
"2302.13971"
] |
2310.02255#201 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Based on the image, Dark Violet does not have the minimum area under the curve. The area under the curve is the highest for the color Green. Multimodal Bard: The correct answer is (B), no. Dark Violet does not have the minimum area under the curve. The color with the minimum area under the curve is Yellow Green. The area under the curve is calculated by summing the areas of all the rectangles under the curve. The rectangle for Dark Violet has a height of 92 and a width of 20, so its area is 1840. The rectangle for Yellow Green has a height of 88 and a width of 40, so its area is 3520. Therefore, the area under the curve for Yellow Green is larger than the area under the curve for Dark Violet. | 2310.02255#200 | 2310.02255#202 | 2310.02255 | [
"2302.13971"
] |
2310.02255#202 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: Based on the scatter plot provided, â area under the curveâ is not represented since there is no curve shown. However, if by â area under the curveâ we mean the overall distribution and position of the points for â Dark Violetâ compared to others, we can analyze the positions of the points. From the scatter plot, the color â Dark Violetâ consistently appears to be the lowest on the y-axis across the entire range of the x-axis. | 2310.02255#201 | 2310.02255#203 | 2310.02255 | [
"2302.13971"
] |
2310.02255#203 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Given the choices: (A) yes (B) no The correct answer is: (A) yes. Correct output: (A) yes Figure 91: The scatter plot presented in this problem is template-generated. While models such as LLaMA-Adapter-V2, LLaVA, and Multimodal Bard encounter difficulties in discerning quantitative relationships between different plot lines, GPT-4V successfully discerns the correct relationships and provides an accurate answer. 87 Published as a conference paper at ICLR 2024 ID: 542 Question: How many models in the figure achieve an Acc score greater than 60? Ground truth: 2 Model Performance GPT3-36 belkand af GPT3-128 at GPT3-1758 E Pane pau LLaMAa2-7B , LLaMA-13B â LLaMA2-138 Fer LLaMA.338 Lbamag- 08 LLaMA-65B Lag. Lamgg-ro8 LLaMA2-70B ungâ opr â LLangg.ta0"â ULang 656 sage? Eg an ange spt cred orgs 16 Pretrain Loss Figure 2: The performance of SFT (blue lines) and ICL (red lines) settings on GSM8K. GPT-4 states they use some part of the GSM8K data in pre-training, and suggest others consider its performance between SFT and ICL. LLaMA-Adapter-V2: There are 10 models in the figure that achieve an Acc score greater than 60.. LLaVA: There are two models in the figure that achieve an Acc score greater than 60. Correct output: 2 | 2310.02255#202 | 2310.02255#204 | 2310.02255 | [
"2302.13971"
] |
2310.02255#204 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Multimodal Bard: There are 8 models in the figure that achieve an Acc score greater than 60. These models are: GPT-4 * GPT3-175B * LLAMA-7B * LLAMA2-7B * LLAMA-13B * LLAMA2-13B * LLAMA-33B * LLAMA-65B * LLAMA2-70B The Acc score is the accuracy of the model, which is the percentage of test cases that the model correctly classifies. An Acc score of 60 means that the model correctly classifies 60% of the test cases. | 2310.02255#203 | 2310.02255#205 | 2310.02255 | [
"2302.13971"
] |
2310.02255#205 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: There are 6 models in the figure that achieve an Acc score greater than 60. Figure 92: The example showcases a scatter plot sourced from an academic paper. Current models, including GPT-4V, struggle to discern the fine-grained relationships depicted in the figure. 88 Published as a conference paper at ICLR 2024 H.4.9 SCIENTIFIC SCENE Answering questions based on scientific scenes poses a challenge in aligning the scientific concepts present in the question text and those in the accompanying figures. GPT-4V demonstrates its superior ability to reason about scientific scenes compared to other models, as evidenced in Figure 1. In the example of Figure 93, GPT-4V adeptly identifies two organisms in the food web and elucidates their relationship. In another instance, shown in Figures 94 and 95, both Multimodal Bard and GPT-4V are able to use knowledge in the physical domain to effectively ground the symbols and variables depicted in the figure. | 2310.02255#204 | 2310.02255#206 | 2310.02255 | [
"2302.13971"
] |
2310.02255#206 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | ID: 685 Question: What would happen to organ- ism c if organism b increased? Choices: (A) decrease (B) increase (C) canâ t predict (D) stay same Ground truth: (B) increase al Organism A Organism B Organism C Organism D (Not drawn to scale) LLaMA-Adapter-V2: If organism b increases, organism c would decrease. This is because organism c is dependent on organism b for its food supply, and an increase in organism b would lead to a decrease in organism c. | 2310.02255#205 | 2310.02255#207 | 2310.02255 | [
"2302.13971"
] |
2310.02255#207 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | LLaVA: If organism b increased, organism c would likely decrease. This is because the growth of organism b would lead to increased competition for resources, such as food and space. As a result, organism c would have to compete more fiercely for these resources, which could lead to a decrease in its population. Multimodal Bard: The answer is (A). If organism B increases, then organism C will decrease. In the food chain, organism B is the predator of organism C. This means that organism B eats organism C. If there are more organism B, then there will be more predation on organism C. This will lead to a decrease in the population of organism C. So the answer is (A). | 2310.02255#206 | 2310.02255#208 | 2310.02255 | [
"2302.13971"
] |
2310.02255#208 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: If Organism B (which appears to be some kind of insect) increased in population, it would mean that there is more available food for Organism C (the fish). Therefore, the population of Organism C would likely: (B) increase. Correct output: (B) increase Figure 93: Models such as LLaMA-Adapter-V2, LLaVA, and Multimodal Bard seem to struggle with discerning the relationships between organisms depicted in the food chain. Instead, GPT-4V recognizes organisms in the food chain and possesses the knowledge to infer their relations. | 2310.02255#207 | 2310.02255#209 | 2310.02255 | [
"2302.13971"
] |
2310.02255#209 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 89 Published as a conference paper at ICLR 2024 ID: 862 Question: One of the most dramatic videos on the web (but entirely fictitious) supposedly shows a man sliding along a long water slide and then being launched into the air to land in a water pool. Letâ s attach some reasonable numbers to such a flight to calculate the velocity with which the man would have hit the water. Figure indicates the launch and landing sites and includes a superimposed coordinate system with its origin conveniently located at the launch site. From the video we take the horizontal flight distance as D = 20.0 m, the flight time as t = 2.50 s, and the launch angle as θ0 = 40.0â | 2310.02255#208 | 2310.02255#210 | 2310.02255 | [
"2302.13971"
] |
2310.02255#210 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | ¦. Find the magnitude of the velocity at launch and at landing. Ground truth: 10.44 Multimodal Bard: Given the following information: * Horizontal flight distance, D = 20.0 m * Flight time, t = 2.50 s * Launch angle, θ0 = 40.0â ¦ We can find the magnitude of the velocity at launch using the following equation: vo = Be + 08y where v0x is the horizontal component of the launch velocity and v0y is the vertical component of the launch velocity. The horizontal component of the launch velocity can be found using the following equation: D t The vertical component of the launch velocity can be found using the following equation: v0y = v0 sin θ0 = 8.00 m/s sin 40.0â ¦ = 5.83 m/s Therefore, the magnitude of the velocity at launch is: # vo = \/ V6 + Voy = V/(8-00 m/s)? + (5.83 m/s)? = 10.44 m/s © 10.4 m/s The magnitude of the velocity at landing will be the same as the magnitude of the velocity at launch, since there is no air resistance in this case. So the answer is 10.4 rounded to two decimal places. Figure 94: This college-level question, derived from the physics curriculum, requires understand- ing the intricate scenario depicted in the scientific figure, possessing domain-specific knowledge, and performing multiple reasoning steps to obtain a conclusive answer. Multimodal Bard correctly processes the reasoning but failed to present the answer in the appropriate format. 90 | 2310.02255#209 | 2310.02255#211 | 2310.02255 | [
"2302.13971"
] |
2310.02255#211 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Published as a conference paper at ICLR 2024 ID: 862 Question: One of the most dramatic videos on the web (but entirely fictitious) supposedly shows a man sliding along a long water slide and then being launched into the air to land in a water pool. Letâ s attach some reasonable numbers to such a flight to calculate the velocity with which the man would have hit the water. Figure indicates the launch and landing sites and includes a superimposed coordinate system with its origin conveniently located at the launch site. From the video we take the horizontal flight distance as D = 20.0 m, the flight time as t = 2.50 s, and the launch angle as θ0 = 40.0â | 2310.02255#210 | 2310.02255#212 | 2310.02255 | [
"2302.13971"
] |
2310.02255#212 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | ¦. Find the magnitude of the velocity at launch and at landing. Ground truth: 10.44 GPT-4V: Given: - D = 20.0 m - t = 2.50 s - θ0 = 40.0â ¦ Using kinematic equations, the horizontal motion is described as: D = V0x à t Where V0x is the horizontal component of the initial velocity. V0x = V0 à cos(θ0) From the first equation: V0 à cos(θ0) = D t Rearranging for V0: V0 = D t à cos(θ0) Substitute the given values to get V0: V0 = 20.0 m 2.50 s à cos(40.0â ¦) Using a calculator: 20.0 2.50 à 0.766 20.0 1.915 V0 = V0 = V0 = 10.44 m/s Thus, the magnitude of the launch velocity (which will be the same as the landing velocity, ignoring air resistance) is 10.44 m/s. Correct output: 10.44 Figure 95: Continuing from Figure 94, GPT-4V generates the correct answer and the associated reasoning steps for this scientific problem. 91 Published as a conference paper at ICLR 2024 H.4.10 SYNTHETIC SCENE Problems involving synthetic scenes require a nuanced understanding of visual objects, such as the numbers, attributes, and positions of these objects, as shown in Figures 96 and 97. Although GPT- 4V demonstrates notable advancements over other models, such as Multimodal Bard, it still falls significantly short of human performance, as shown in Figure 1. | 2310.02255#211 | 2310.02255#213 | 2310.02255 | [
"2302.13971"
] |
2310.02255#213 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | ID: 52 Question: Subtract all big red metallic spheres. all big brown matte things. How many objects are left? Ground truth: 6 Subtract LLaMA-Adapter-V2: The final value is 1, as there is only one object left after subtracting all big red metallic spheres and all big brown matte things. LLaVA: After subtracting all big red metallic spheres and big brown matte things, there are 3 objects left. Multimodal Bard: Given the image, there are 4 objects: * 1 big red metallic sphere * 1 big brown matte cube * 1 small red metallic sphere * 1 small brown matte sphere If we subtract all big red metallic spheres, there will be 3 objects left: * 1 big brown matte cube * 1 small red metallic sphere * 1 small brown matte sphere If we subtract all big brown matte things, there will be 2 objects left: * 1 small red metallic sphere * 1 small brown matte sphere Therefore, if we subtract all big red metallic spheres and big brown matte things, there will be 2 objects left. | 2310.02255#212 | 2310.02255#214 | 2310.02255 | [
"2302.13971"
] |
2310.02255#214 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: There are no big red metallic spheres in the image. There is 1 big brown matte object (the cube). Total objects to be subtracted = 0 + 1 = 1 Total objects in the image = 7 Objects left = 7 - 1 = 6 There are 6 objects left. Correct output: 6 Figure 96: In this example, GPT-4V is able to recognize the objects with specific attributes from the synthetic figure and then perform arithmetic reasoning to get the correct answer. 92 Published as a conference paper at ICLR 2024 ID: 177 Question: | 2310.02255#213 | 2310.02255#215 | 2310.02255 | [
"2302.13971"
] |
2310.02255#215 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Is the number of brown tandem bikes that are to the left of the small blue matte car greater than the number of tiny blue biplanes? Choices: (A) Yes (B) No Ground truth: (A) Yes LLaMA-Adapter-V2: Yes, the number of brown tandem bikes that are to the left of the small blue matte car is greater than the number of tiny blue biplanes. Correct output: (A) Yes LLaVA: The correct answer is (B) No. | 2310.02255#214 | 2310.02255#216 | 2310.02255 | [
"2302.13971"
] |
2310.02255#216 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | There are more tiny blue biplanes to the left of the small blue matte car than brown tandem bikes. Multimodal Bard: The answer is (B). There are two brown tandem bikes to the left of the small blue matte car, and there is only one tiny blue biplane. So the number of brown tandem bikes is not greater than the number of tiny blue biplanes. GPT-4V: Looking at the image, there is 1 brown tandem bike to the left of the small blue matte car. There are no tiny blue biplanes in the image. Thus, the number of brown tandem bikes that are to the left of the small blue matte car is (A) Yes greater than the number of tiny blue biplanes. Correct output: (A) Yes Figure 97: The question in this example requires the models to accurately extract the attributes of different objects from the synthetic figure. GPT-4V infers the correct answer based on its generated detailed description of the figure. | 2310.02255#215 | 2310.02255#217 | 2310.02255 | [
"2302.13971"
] |
2310.02255#217 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 93 Published as a conference paper at ICLR 2024 # H.4.11 TABLE Tables serve as a powerful tool to present and summarize large amounts of data in a comprehensible manner. In particular, GPT-4V has shown significant advancements over other foundation models and even surpasses human performance on table-related reasoning tasks, as shown in Figure 1. The example in Figure 98 shows a complex table taken from an academic paper. GPT-4V can accurately pinpoint the target cells among numerous rows and columns. Figure 99 shows a QA task in which the answer needs to be derived from the table regarding the push-up competition. GPT-4V is the only model that can produce the correct answer. ID: 923 Choices: (A) Transformers Agent (GPT-4) (B) LLaMA-Adapter v2-7B (C) LLaVA-7B (D) Otter-9B (E) MM-ReAct-GPT-3.5 (F) LLaVA-13B (LLaMA-2) (G) MM-ReAct-GPT-4 Ground truth: (F) LLaVA-13B (LLaMA-2) LLaMA-Adapter-V2: The correct answer is (G) MM-ReAct-GPT-4. LLaVA: The correct answer is (B) LLaMA-Adapter v2-7B, which achieves the highest score in terms of Rec. # Multimodal Bard: The answer is (G) MM-ReAct-GPT-4. MM-ReAct-GPT-4 achieves the highest score of 33.1 in terms of Rec, which is the ability to recognize and understand text. The other models have the following scores in terms of Rec: * Transformers Agent (GPT-4): 18.2 * LLaMA-Adapter v2-7B: 16.8 * LLaVA-7B: 28.0 * Otter-9B: 28.4 * MM-ReAct-GPT-3.5: 24.2 * LLaVA-13B (LLaMA-2): 32.9 GPT-4V: The model that achieves the highest score in terms of Rec is â LLaVA-13B (LLaMA-2)â with a score of 39.2%. | 2310.02255#216 | 2310.02255#218 | 2310.02255 | [
"2302.13971"
] |
2310.02255#218 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | The correct option is: (F) LLaVA-13B (LLaMA-2) Correct output: (F) LLaVA-13B (LLaMA-2) Figure 98: The problem presents a complex table from an academic paper, complete with a caption and multiple rows and columns. This format poses challenges to models like LLaMA-Adapter-V2, LLaVA, and Multimodal Bard. However, GPT-4V demonstrates its exceptional reasoning abilities within tabular contexts. | 2310.02255#217 | 2310.02255#219 | 2310.02255 | [
"2302.13971"
] |
2310.02255#219 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 94 Published as a conference paper at ICLR 2024 Push-up competition (number of push-ups) Stem | Leaf 1335778 ID: 265 Question: Justineâ s P.E. class participated in a push-up competition, and Justine wrote down how many push-ups each person could do. How many people did at least 60 push-ups? (Unit: people) Ground truth: 11 LLaMA-Adapter-V2: Justineâ s P.E. class participated in a push-up competition, and Justine wrote down how many push-ups each person could do. There were 8 people in the competition, and at least 60 push-ups were done by 3 people. | 2310.02255#218 | 2310.02255#220 | 2310.02255 | [
"2302.13971"
] |
2310.02255#220 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | The final value is 3. LLaVA: In the image, there are a total of 13 people participating in the push-up competition. Among them, there are two people who did at least 60 push-ups. Multimodal Bard: There are 4 people who did at least 60 push-ups. The person with the highest number of push-ups did 778, so the only people who did at least 60 push-ups are the ones who did 60, 77, 78, and 778. There is one person who did 60 push-ups, one person who did 77 push-ups, one person who did 78 push-ups, and one person who did 778 push-ups. So there are a total of 1 + 1 + 1 + 1 = 4 people who did at least 60 push-ups. | 2310.02255#219 | 2310.02255#221 | 2310.02255 | [
"2302.13971"
] |
2310.02255#221 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: For the stem-and-leaf plot provided: 6 represents the tens value, and the digits 3, 6, 7, and 7 represent the ones value. So, the number of push-ups represented by the 6 row are 63, 66, 67, and 67. 7 represents the tens value, and the digits 1, 3, 3, 5, 7, 7, and 8 represent the ones value. So, the number of push-ups represented by the 7 row are 71, 73, 73, 75, 77, 77, and 78. Counting these numbers: 4 from the 6 row + 7 from the 7 row = 11. Correct output: 11 Figure 99: This figure features a stem-and-leaf plot used for statistical analysis. While LLaMA- Adapter-V2, LLaVA, and Multimodal Bard struggle to comprehend this domain-specific table, GPT- 4V excels in understanding its content. 95 Published as a conference paper at ICLR 2024 H.4.12 OTHER VISUAL CONTEXTS On the reasoning tasks using other visual contexts, GPT-4V achieves a higher overall accuracy than all the other models, as depicted in Figure 1. For instance, GPT-4V is the only model that is capable of generating the correct answer to the question regarding a violin plot, as shown in Figure 100. Audio-Audio Similarity Text-Audio Similarity ) N N MusicLDM (mix-up) MusicLDM (original) MusicLDM (BLM) MusicLDM (BAM) MuBERT Figure 3: The violin plot of the audio-audio similarity, and the text-to-audio similarity. | 2310.02255#220 | 2310.02255#222 | 2310.02255 | [
"2302.13971"
] |
2310.02255#222 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | ID: 895 Question: Which model has the lowest Audio-Audio Similarity and Text-Audio Similarity scores overall? Choices: (A) MusicLDM (mix-up) (B) MusicLDM (original) (C) MusicLDM (BLM) (D) MusicLDM (BAM) (E) MuBERT Ground truth: (E) MuBERT LLaMA-Adapter-V2: The model with the lowest Audio-Audio Similarity and Text-Audio Similarity scores overall is MusicLDM (BLM). LLaVA: Based on the image, the model with the lowest Audio-Audio Similarity and Text-Audio Similarity scores overall is MusicLDM (mix-up). Multimodal Bard: | 2310.02255#221 | 2310.02255#223 | 2310.02255 | [
"2302.13971"
] |
2310.02255#223 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | The answer is (C). MusicLDM (BLM) has the lowest Audio-Audio Similarity and Text-Audio Similarity scores overall. The violin plot shows that the distribution of audio-audio similarity scores for MusicLDM (BLM) is the most spread out, indicating that there is a lot of variation in the scores. The distribution of text-audio similarity scores for MusicLDM (BLM) is also the most spread out. This suggests that MusicLDM (BLM) is not very consistent in its ability to capture the similarity between audio and text. | 2310.02255#222 | 2310.02255#224 | 2310.02255 | [
"2302.13971"
] |
2310.02255#224 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V: The model with the lowest Audio-Audio Similarity and Text-Audio Similarity scores overall is MuBERT. # Correct output: (E) MuBERT Figure 100: Violin plots are widely used in academic papers to visualize the distribution of data and probability density. In this example, only GPT-4V is able to generate the correct output from the figure. 96 Published as a conference paper at ICLR 2024 H.5 SELF-VERIFICATION IN GPT-4V Self-verification is a social psychological theory asserting that people desire others to perceive them as they see themselves. Consequently, individuals will take active measures to ensure that others view them in ways that confirm their stable self-concepts (Talaifar & Swann, 2020). Interestingly, in our experiments, GPT-4V demonstrates an ability similar to self-verification. The model can inspect its own behaviors during the course of reasoning and can take active actions to correct its mistakes. Note that self-verification we discuss here differs from several recent works on improving the modelâ s outputs based on external feedback (Peng et al., 2023) or additional gener- ations (Yang et al., 2023b). The examples in Figures 101 and 103 show that GPT-4V, on its own, can inspect a set of candidate answers and identify the one that is valid and meets all the given con- straints. The multi-step reasoning example in Figure 102 shows that GPT-4V can verify the validity of (the result of) each reasoning step, and explore alternative approaches if any invalid (intermediate) result is detected (e.g., a negative value for length). Although self-verification does not guarantee an accurate response even after multiple tries, espe- cially when applying GPT-4V to visual perception or mathematical reasoning in intricate scenarios (see Figure 104), it is instrumental in improving the model performance on MATHVISTA. We also found that GPT-4Vâ s self-verification is weaker for non-English tasks, such as Mandarin, as shown in Figure 105. It is also worth noting that self-verification does not emerge in other foundation mod- els we studied, or at least it is not as robust as that of GPT-4V. | 2310.02255#223 | 2310.02255#225 | 2310.02255 | [
"2302.13971"
] |
2310.02255#225 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | As shown in Figure 106, Multimodal Bard first attempts a natural language solution, followed by a program-assisted one for verification. However, the program-aided solution leads to a different and incorrect prediction. The emergent ability of self-verification highlights GPT-4Vâ s potential in solving rigorous reasoning and theorem-proving tasks. One of the most exciting research topics for future work is to develop a mechanism that allows the model to activate self-verification consistently at the right time and to use a set of alternative approaches that maximize the success rate of task completion. ID: 615 Question: | 2310.02255#224 | 2310.02255#226 | 2310.02255 | [
"2302.13971"
] |
2310.02255#226 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | If two sides of a triangle measure 12 and 7, which of the following cannot be the perimeter of the triangle? Choices: (A) 29 (B) 34 (C) 37 (D) 38 Ground truth: (D) 38 GPT-4V: The sum of the lengths of any two sides of a triangle must be greater than the length of the third side. Let x be the length of the third side. From the triangle inequality: 12 + 7 > x => x < 19 Similarly: 7 + x > 12 => x > 5 And: 12 + x > 7 => x > â 5 (This will always be true since the length cannot be negative) So, 5 < x < 19 Now, letâ s check the given options: (A) 29 => x = 29 â 12 â 7 = 10 (Possible) (B) 34 => x = 34 â 12 â 7 = 15 (Possible) (C) 37 => x = 37 â 12 â 7 = 18 (Possible) (D) 38 => x = 38 â 12 â 7 = 19 (Not Possible) Correct output: (D) 38 Figure 101: Given several possible values for the length of the third side, GPT-4V verifies each candidate answer to ensure that it meets the given constraints (highlighted in blue). 97 Published as a conference paper at ICLR 2024 ID: 6 Question: As shown in the figure, if CB = 4.0, DB = 7.0, and D is the midpoint of AC, then the length of AC is (). Choices: (A) 6 cm (B) 7 cm (C) 4 cm (D) 5 cm Ground truth: (A) 6 cm GPT-4V: From the figure, we can deduce: AD = DC (since D is the midpoint of AC) Now, using the entire line segment AB: AB = AD + DB AB = DC + DB Given DB = 7.0 cm, we can rearrange to find: AD (or DC) = AB â DB AD = AB â 7.0 Now, letâ s find AB: From the given information, AB = AC + CB AB = AC + 4.0 Using the previously deduced equation: | 2310.02255#225 | 2310.02255#227 | 2310.02255 | [
"2302.13971"
] |
2310.02255#227 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | AD = AC + 4.0 â 7.0 AD = AC â 3.0 Since D is the midpoint of AC, AD will be half the length of AC: AC = 2(AD) From the above: AC â 3.0 = 2(AD) Given that AD and DC are equal, we can represent DC as: DC = AC/2 Using the relation CB = 4.0 cm and DB = 7.0 cm: AC/2 + 7.0 = 4.0 AC/2 = â 3.0 This is not possible as lengths cannot be negative. Letâ s re-evaluate our approach. From the original figure: AB = AD + DB AB = DC + DB Given DB = 7.0 cm, we have: AD = AB â 7.0 AD = AC/2 â 7.0 Using the relation AB = AC + CB: AC/2 + 7.0 = AC + 4.0 Subtracting AC/2 from both sides: 7.0 = AC/2 + 4.0 Subtracting 4 from both sides: AC/2 = 3.0 Multiplying both sides by 2: AC = 6.0 cm # Correct output: (A) 6 cm Figure 102: GPT-4V generates the result as AC/2 = â 3.0. It verifies the validity of the value and concludes it cannot be negative. The model then re-evaluates its approach and adopts a new method to obtain the correct answer, highlighted in blue. 98 | 2310.02255#226 | 2310.02255#228 | 2310.02255 | [
"2302.13971"
] |
2310.02255#228 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Published as a conference paper at ICLR 2024 ID: 28 Question: AB is a diameter, AC = 8 inches, and BC = 15 inches. Find the radius of the circle. Choices: (A) 7.5 (B) 8 (C) 8.5 (D) 17 Ground truth: (C) 8.5 GPT-4V: Given that AB is a diameter and AC = 8 inches and BC = 15 inches, we can use the Pythagorean theorem to determine the length of AB. The right triangle is formed by A, the center of the circle, and C makes AC the shorter leg, the radius the longer leg, and BC the hypotenuse. Let r be the radius of the circle. Using the Pythagorean theorem: AC 2 + r2 = BC 2 Substituting the given lengths: 82 + r2 = 152 64 + r2 = 225 r2 = 161 Taking the square root of both sides: r = â 161 However, r is not any of the choices. We need to find the diameter, which is twice the radius. AB = 2r = 2 â 161 But, since the diameter is just twice the radius and the radius is one of the choices, the radius is: r = â 161/2 = 8.5 (approximately) # Correct output: (C) 8.5 â 161, GPT-4V performs two â 161/2. It then selects the Figure 103: Starting by computing the initial radius length as r = verifications resulting in different answers: r = 2 161 and r = matching option from the choices as its final prediction, highlighted in blue. â 99 | 2310.02255#227 | 2310.02255#229 | 2310.02255 | [
"2302.13971"
] |
2310.02255#229 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Published as a conference paper at ICLR 2024 ID: 644 Question: What is the missing number in the picture? Choices: (A) 6 (B) 8 (C) 10 (D) 11 Ground truth: (B) 8 Solution: 8 + 4 â 2 = 10, 3 + 2 â 3 = 2, 12 + 1 â 5 = 8 GPT-4V: Letâ s analyze the pattern in the given triangles: 1. For the first triangle: (8 â 4) Ã 2 = 8 2. For the second triangle: (3 â 2) Ã 3 = 3 Applying the same pattern to the third triangle: (12 â 1) Ã 5 = 55 However, since the answer is in single digits, it seems we take only the unitâ s place of the result. So, the missing number is: 5 (unitâ s place of 55) This is not in the given options. Therefore, letâ s explore another possible pattern: 1. For the first triangle: (8 + 4)/2 = 6 2. For the second triangle: (3 + 2)/3 = 5/3 = 1.67 (This does not match the number inside) But letâ s continue with this pattern for the third triangle: (12 + 1)/5 = 13/5 = 2.6 Again, this does not match any of the options. Considering other possible operations: 1. For the first triangle: (8 Ã 4)/2 = 16 2. For the second triangle: (3 Ã 2)/3 = 2 3. For the third triangle: (12 Ã 1)/5 = 12/5 = 2.4 Again, this does not match any of the options. Given the information and the options provided, I am unable to find a consistent pattern that applies to all the triangles. Therefore, I cannot determine the missing number from the given choices. Figure 104: In this logical reasoning question, GPT-4V verifies its generated solution and retries twice. However, it still fails to produce an answer that matches any of the options. Consequently, it claims that it cannot determine the missing number from the given choices (highlighted in blue). | 2310.02255#228 | 2310.02255#230 | 2310.02255 | [
"2302.13971"
] |
2310.02255#230 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 100 Published as a conference paper at ICLR 2024 ID: 32 Question: å¦ å ¾ ï¼ å ¨ ABCD ä¸ ï¼ AB = ACï¼ â CAB = 40â ¦ï¼ å â D ç åº¦æ °æ ¯ï¼ ï¼ ã (In the figure, within ABCD, AB = AC, and â CAB = 40â ¦, then the measure of â D is ().) Choices: (A) 40° (B) 50° (C) 60° (D) 70° Ground truth: (D) 70° GPT-4V: ç» å® AB = ACï¼ â ³ABC æ ¯ç è °ä¸ è§ å½¢ã ç ±äº â CAB = 40â ¦ï¼ æ 以 â CBA ä¹ ä¸º 40â ¦ ï¼ å ä¸ºå ¨ç è °ä¸ è§ å½¢ä¸ ï¼ ä¸¤ä¸ªåº è§ æ ¯ç ¸ç ç ï¼ ã (Given AB = AC, â ³ABC is an isosceles triangle. Since â CAB = 40â ¦, â CBA is also 40â ¦ (because in an isosceles triangle, the two base angles are equal).) è è å 边形 ABCDï¼ å ¶å è§ å 为 360â ¦ã å·²ç ¥ â CAB = 40â ¦ï¼ â CBA = 40â ¦ï¼ å¹¶ä¸ â BCA å â D æ ¯è¡¥è§ ã (Considering the quadrilateral ABCD, its interior angles sum up to 360â ¦. We know that â CAB = 40â ¦, â CBA = 40â ¦, and â BCA and â D are supplementary.) æ ä»¬å ¯ä»¥æ ¾å ° â BCA ç è§ åº¦ä¸º: (We can find the angle of â BCA as:) â BCA = 180â ¦ â 40â ¦ â 40â ¦ = 100â ¦ ç ±äº â BCA å â D æ | 2310.02255#229 | 2310.02255#231 | 2310.02255 | [
"2302.13971"
] |
2310.02255#231 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | ¯è¡¥è§ ï¼ æ 们æ : (As â BCA and â D are supplementary, we have:) â D = 180â ¦ â 100â ¦ = 80â ¦ ä½ è¿ ä¸ æ ¯ç» å® ç é 项ã æ ä»¥ï¼ è®©æ 们é æ °æ£ æ ¥ã (But this isnâ t a given option. So, letâ s recheck.) 注æ å ° â BCA æ ¯ä¸ ä¸ªå¤ è§ ï¼ å® ç äº ä¸¤ä¸ªé ç ¸é »ç å è§ ä¹ å ã æ 以: (Notice that â BCA is an exterior angle, which equals the sum of two non-adjacent interior angles. So:) â BCA = â CAB + â CBA â BCA = 40â ¦ + 40â ¦ = 80â ¦ æ 以 â D = 180â ¦ â 80â ¦ = 100â ¦ã ä½ è¿ ä¹ ä¸ æ ¯ä¸ ä¸ªé 项ã (Thus, â D = 180â ¦ â 80â ¦ = 100â ¦. But this isnâ t an option either.) ç ä¸ ä¸ ï¼ æ 好å è è é äº ã (Wait, I seem to have thought it wrong.) â BCA å â BDA æ ¯è¡¥è§ ï¼ å 为 ABCD æ ¯ä¸ ä¸ªå è¾¹å½¢ï¼ æ ä»¥å® ä»¬ä¸¤ä¸ªç å 为 180â ¦ã (â BCA and â BDA are supplementary because ABCD is a quadrilateral, so their sum is 180â ¦.) å æ ¤, â BDA = 180â ¦ â 80â ¦ = 100â ¦ï¼ ä½ æ ¯ â D æ ¯ â BDA å â CBA ç å·®ã (Therefore, â BDA = 180â ¦ â 80â ¦ = 100â ¦, but â D is the difference between â BDA and â CBA.) # æ 以 (So) | 2310.02255#230 | 2310.02255#232 | 2310.02255 | [
"2302.13971"
] |
2310.02255#232 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | â D = 100â ¦ â 40â ¦ = 60â ¦ Figure 105: The question is presented in Chinese, with the translated English provided in paren- theses after the Chinese text. GPT-4V is capable of performing self-verification until it produces a solution that matches one of the given options (highlighted in blue). 101 Published as a conference paper at ICLR 2024 ID: 44 Question: Chase wants to buy 4 kilograms of oval beads and 5 kilograms of star-shaped beads. How much will he spend? (Unit: $) Ground truth: 18 Solution: 8 + 4 â 2 = 10, 3 + 2 â 3 = 2, 12 + 1 â 5 = 8 | 2310.02255#231 | 2310.02255#233 | 2310.02255 | [
"2302.13971"
] |
2310.02255#233 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Multimodal Bard: The cost of 4 kilograms of oval beads is 4 â 2 = $8. The cost of 5 kilograms of star-shaped beads is 5 â 2 = $10. So the total cost is $8 + $10 = $18. Therefore, Chase will spend $18. I also wrote a Python code to calculate the total cost and got the same answer of 35. def calculate_cost(oval_beads, star_shaped_beads): oval_bead_cost = 2 star_shaped_bead_cost = 2 total_cost = oval_beads * oval_bead_cost + star_shaped_beads * star_shaped_bead_cost return total_cost print(calculate_cost(4, 5)) # 35 | 2310.02255#232 | 2310.02255#234 | 2310.02255 | [
"2302.13971"
] |
2310.02255#234 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Figure 106: Multimodal Bard seems incapable of verifying its solutions and revisiting them if found invalid. For instance, in this example, after initially producing a natural language solution, Mul- timodal Bard subsequently generates Python code. However, the executed result of the code (35) differs from its initial solution (18). 102 Published as a conference paper at ICLR 2024 # H.6 SELF-CONSISTENCY FOR GPT-4V Self-consistency (Wang et al., 2022) is a decoding strategy for chain-of-thought prompting (Wei et al., 2022b). A diverse set of reasoning paths is sampled, and the most consistent answer is se- lected as the final prediction. Moving beyond vanilla greedy decoding, this method resorts to the inherent coherence and reliability of multiple reasoning trajectories to produce a more trustworthy conclusion. Self-consistency has been widely employed in LLMs for complex reasoning tasks, such as math word problems and commonsense reasoning. In our experiments, we validated the effectiveness of using self-consistency for GPT-4V on MATH- VISTA. Given a question and context, we ran GPT-4V multiple times to obtain a set of different reasoning paths and then selected the most frequent answer as the final prediction. We found that self-consistency is instrumental in rectifying visual perception errors (Figure 107), correcting cal- culation mistakes (Figure 108), and mitigating hallucinations (Figure 109). In comparison, self- consistency is less effective when GPT-4V has difficulties in interpreting complex visual contexts (Figures 110, 111) or extracting salient information from images (Figure 112). | 2310.02255#233 | 2310.02255#235 | 2310.02255 | [
"2302.13971"
] |
2310.02255#235 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | ID: 648 Question: What this function? Ground truth: -1 is the minimum value of 05 00 05 10 1.0 GPT-4V (Reasoning Path 1): The minimum value of the function is observed where the graph is at its lowest point along the y-axis. From the graph, the minimum value lies on the y-axis between -0.5 and 0. Based on the visual representa- tion, it appears that the minimum value of y is: -0.5. So, the answer is -0.5. Wrong output: -0.5 | 2310.02255#234 | 2310.02255#236 | 2310.02255 | [
"2302.13971"
] |
2310.02255#236 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V (Reasoning Path 2): The graph shown represents a function with circular data points plotted across the x-axis. To determine the minimum value of the function, we need to locate the lowest point on the graph. From the graph, it is evident that the lowest point (or the minimum value of the function) is at a y-coordinate of approximately -1. Therefore, the minimum value of the function is -1. Correct output: -1 GPT-4V (Reasoning Path 3): The graph depicts the function using circular points. The minimum value of the function can be observed by identifying the lowest point on the graph. From the graph, the lowest point seems to be at y = -1. Therefore, the minimum value of this function is -1. Correct output: -1 Figure 107: In this algebraic reasoning task, GPT-4V displays inconsistency in recognizing the value range of the function. After reevaluating GPT-4V three times, three distinct reasoning paths emerge. Utilizing self-consistency minimizes variation in the generated results, converging on the most frequent prediction (-1) as the final answer. | 2310.02255#235 | 2310.02255#237 | 2310.02255 | [
"2302.13971"
] |
2310.02255#237 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 103 Published as a conference paper at ICLR 2024 ID: 331 Question: Several people compared how many Web pages they had visited. What is the mean of the numbers? Ground truth: 64 GPT-4V (Reasoning Path 1): To find the mean of the numbers, you add up all the numbers and then divide by the quantity of numbers. Mean = (84 + 56 + 70 + 68 + 60 + 47 + 75 + 52) ÷ 8 = 462 ÷ 8 = 57.75 However, since the question requires an integer answer, the mean number of web pages visited is: 58. Wrong output: 58 GPT-4V (Reasoning Path 2): To find the mean number of web pages visited: First, sum the number of web pages for all the individuals: Bonnie: 84, Winston: 56, Percy: 70, Derek: 68, Jake: 60, Finn: 47, Carter: 75, Kurt: 52 Total = 84 + 56 + 70 + 68 + 60 + 47 + 75 + 52 = 512 Next, divide by the number of individuals (8): Mean = 512 ÷ 8 = 64 The mean number of web pages visited is 64. Correct output: 64 GPT-4V (Reasoning Path 3): | 2310.02255#236 | 2310.02255#238 | 2310.02255 | [
"2302.13971"
] |
2310.02255#238 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | First, Iâ ll list the number of web pages each person visited: Bonnie: 84, Winston: 56, Percy: 70, Derek: 68, Jake: 60, Finn: 47, Carter: 75, Kurt: 52 Next, Iâ ll group some numbers together for easier addition: (84 + 56) = 140 (70 + 68) = 138 (60 + 47) = 107 (75 + 52) = 127 Now, Iâ ll sum these results: 140 + 138 + 107 + 127 = 512 Now, Iâ ll divide the sum by the number of individuals to get the mean: 512 ÷ 8 = 64 So, the mean number of web pages visited is 64. Correct output: 64 Figure 108: For this arithmetic reasoning question, GPT-4V arrives at the answers of 58, 64, and 64 through three distinct reasoning paths. Applying self-consistency would confirm the correct answer to be 64. 104 | 2310.02255#237 | 2310.02255#239 | 2310.02255 | [
"2302.13971"
] |
2310.02255#239 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Published as a conference paper at ICLR 2024 ID: 430 Question: The derivative of f (x) at x = 0 is that at x = 5 Choices: (A) larger than (B) equal to (C) smaller than Ground truth: (C) smaller than foo) = [2x - 3|+41 GPT-4V (Reasoning Path 1): The graph represents the function f (x) = |2x â 3| + 1. The slope (or the derivative) of f (x) is represented by the inclination of the line. At x = 0, the slope is negative (the line is declining), and at x = 5, the slope is positive (the line is inclining). The magnitude of the slopes appears to be the same; only their directions differ. However, when considering the magnitude of the derivative, the value at x = 0 is negative, and the value at x = 5 is positive, but they have the same absolute value. Wrong output: (B) equal to GPT-4V (Reasoning Path 2): The given function is f (x) = |2x â 3| + 1. The absolute value function is piecewise defined, so its derivative will have different values depending on where you take it. For 2x â 3 > 0, which is x > 3 2 or 1.5, f (x) = 2x â 3. The derivative is f â ²(x) = 2. For 2x â 3 < 0, which is x < 3 2 or 1.5, f (x) = â 2x + 3. The derivative is f â ²(x) = â 2. At x = 0, the derivative f â ²(x) = â 2. At x = 5, the derivative f â ²(x) = 2. Therefore, the derivative of f (x) at x = 0 is smaller than that at x = 5. Correct output: (C) smaller than GPT-4V (Reasoning Path 3): The function provided is f (x) = |2x â 3| + 1. The absolute value function has a V-shaped graph, as shown in the image. | 2310.02255#238 | 2310.02255#240 | 2310.02255 | [
"2302.13971"
] |
2310.02255#240 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | The slope (or derivative) of the line to the left of the vertex is negative, and the slope of the line to the right of the vertex is positive. The derivative of f (x) at x = 0 corresponds to the slope of the line on the left side of the vertex, which is negative (downward slope). The derivative of f (x) at x = 5 corresponds to the slope of the line on the right side of the vertex, which is positive (upward slope). Since a positive value is always larger than a negative value, the derivative of f (x) at x = 0 is smaller than that at x = 5. Correct output: (C) smaller than Figure 109: This is another instance where the application of self-consistency across three reasoning paths results in a correct final prediction. 105 Published as a conference paper at ICLR 2024 ID: 157 Question: What number is shown? Ground truth: 9335 CC CC CC GPT-4V (Reasoning Path 1): Letâ s break down the number shown using base-ten blocks: There are 6 large cubes, each representing 1, 000. So, 6 Ã 1, 000 = 6, 000. - There are 7 flats, each representing 100. So, 7 Ã 100 = 700. - There are 4 rods, each representing 10. So, 4 Ã 10 = 40. - There are 3 units, each representing 1. So, 3 Ã 1 = 3. Adding all of these values together: 6, 000 + 700 + 40 + 3 = 6, 743 The number shown is 6, 743. Wrong output: 6743 GPT-4V (Reasoning Path 2): The image displays a series of cubes that appear to represent numbers in base 10 using place value. From left to right: 1. | 2310.02255#239 | 2310.02255#241 | 2310.02255 | [
"2302.13971"
] |
2310.02255#241 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | There are six large cubes, which represent a value of 6 Ã 1000 = 6000. 2. There are three medium-sized cubes, which represent a value of 3Ã 100 = 300. 3. There are two small cubes, which represent a value of 2 Ã 10 = 20. 4. The vertical stack of single unit cubes shows nine cubes, which represent a value of 9 Ã 1 = 9. Adding up these values: 6000 + 300 + 20 + 9 = 6329 The number shown is 6329. Wrong output: 6329 GPT-4V (Reasoning Path 3): The image displays a series of cubes representing numbers in base 10 using place value. From left to right: 1. There are seven large cubes, which represent a value of 7 Ã 1000 = 7000. 2. There are two medium-sized cubes, which represent a value of 2 Ã 100 = 200. 3. There are nine small cubes, which represent a value of 9 Ã 10 = 90. 4. The vertical stack of single unit cubes shows eight cubes, which represent a value of 8 Ã 1 = 8. Adding up these values: 7000 + 200 + 90 + 8 = 7298 The number shown is 7298. Wrong output: 7298 Figure 110: In this arithmetic reasoning task, GPT-4V struggles with comprehending the abstract scene, leading to its inability to calculate the correct answer through multiplication. Multiple at- tempts yield varying incorrect solutions. Consequently, utilizing self-consistency does not verify the definitive right answer. | 2310.02255#240 | 2310.02255#242 | 2310.02255 | [
"2302.13971"
] |
2310.02255#242 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 106 Published as a conference paper at ICLR 2024 ID: 551 Question: What is the difference be- tween the highest tattoos in male and the least in female? Ground truth: 14 (%) Solution: 27% (male in 2015) - 13% (female in 2008) = 14% 2008 @ Mele @ Female GPT-4V (Reasoning Path 1): The highest percentage of tattoos in males is 31% (in 2015), and the least percentage of tattoos in females is 15% (in 2008). # Difference = 31% - 15% = 16% Wrong output: 16% | 2310.02255#241 | 2310.02255#243 | 2310.02255 | [
"2302.13971"
] |
2310.02255#243 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | GPT-4V (Reasoning Path 2): The highest percentage of males with at least one tattoo is in 2015 at 31%. The lowest percentage of females with at least one tattoo is in 2003, 2008, and 2012, all at 15%. The difference between these values is: 31% - 15% = 16% Wrong output: 16% GPT-4V (Reasoning Path 3): To determine the difference between the highest tattoos in males and the least in females: Highest tattoos in males: 31% (in 2015) Least tattoos in females: 15% (in 2003 and 2008) Difference: 31% - 15% = 16% Wrong output: 16% | 2310.02255#242 | 2310.02255#244 | 2310.02255 | [
"2302.13971"
] |
2310.02255#244 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Prompt: What is the percentage of tattoos in females in 2008? GPT-4V: The percentage of tattoos in females in 2008 is 15%. Prompt: What is the percentage of tattoos in males in 2015? GPT-4V: The percentage of tattoos in males in 2015 is 31%. @ Mate @ Female 2008 Figure 111: In this instance, GPT-4V persistently produces incorrect predictions across multiple evaluations. When prompted to detail the essential elements of the image, GPT-4V provides inac- curate descriptions. This suggests that if GPT-4V misinterprets vital visual cues, relying on self- consistency may not necessarily guide it towards the correct conclusion. | 2310.02255#243 | 2310.02255#245 | 2310.02255 | [
"2302.13971"
] |
2310.02255#245 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 107 Published as a conference paper at ICLR 2024 ID: 965 Question: Whatâ s the lowest number yard line that you can see? Ground truth: 30 f ; GPT-4V (Reasoning Path 1): The lowest number yard line visible in the image is the 10-yard line. GPT-4V (Reasoning Path 2): From the image provided, the lowest number yard line that is clearly visible is the â 10â yard line. GPT-4V (Reasoning Path 3): From the image provided, the lowest number yard line that can be seen is the 10-yard line. GPT-4V (Reasoning Path 4): From the image, the lowest number yard line that is visible is the â 1â yard line. GPT-4V (Reasoning Path 5): From the image, the lowest number yard line that is visible is the 20-yard line. Prompt: List the yard line numbers that are visible in the image. GPT-4V: The visible yard line numbers in the image are 10, 20, 30. Prompt: Can you see the yard numbers 40 and 50 in the image? GPT-4V: No, the yard numbers 40 and 50 are not visible in the image. Figure 112: This example demonstrates that if GPT-4V misinterprets key information from an im- age, the application of self-consistency may not aid in producing the correct final prediction. We evaluated GPT-4V five times for the question â | 2310.02255#244 | 2310.02255#246 | 2310.02255 | [
"2302.13971"
] |
2310.02255#246 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Whatâ s the lowest number yard line that you can see?â and received consistently incorrect responses. Further, when prompted to describe the yard line numbers present in the image, GPT-4V provided incorrect descriptions, indicating its difficulty in accurately recognizing them. 108 Published as a conference paper at ICLR 2024 GPT-4V FOR MULTI-TURN HUMAN-AI INTERACTION section investigates the use of GPT-4V for multi-turn human-AlI interaction on MATHVISTA, exemplified in the goal-directed dialog in Figure 113. that GPT-4V is effective in engaging multi-turn goal-directed conversations with users. particular, GPT-4V can make good use of hints (e.g., user feedback or responses) to guide the to generate desirable results. For instance, it can (1) rectify visual perception errors on hints (Figure 114), (2) reassess reasoning steps and calculations (Figure 115), (3) correct misinformation using user-provided domain-specific knowledge (Figure 116), and (4) aggregate contexts over multiple turns in a human-AI conversation (Figures 117 and 118). observed failure cases in our evaluation. For instance, GPT-4V struggles to generate correct when questions and user hints are ambiguous (Figure 119), or when the model fails abstract shapes and concepts visually (Figure 120). These failures motivate the develop- more powerful, conversational foundation models. sour apple candy $0.06 piece of gum $0.07 gummy worm $0.09 lemon drop $0.05 piece of licorice $0.07 butterscotch candy $0.07 How much money does Luca need to buy a sour apple candy and a butterscotch candy? (Unit: $) S The total cost for a sour apple candy and a butterscotch candy is $0.13. Luca would need at least $0.13 to purchase both items. yr @ butterscotch candy. Adding the two together gives us a total cost of $0.13. Hence, Luca would need at least $0.13 to purchase both items. | 2310.02255#245 | 2310.02255#247 | 2310.02255 | [
"2302.13971"
] |
2310.02255#247 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | S . The cost for each item is $0.06 for the sour apple candy and $0.07 for the H.7 # This # as We found In conversion based # intricate # We also responses understand # ment of # Figure 113: A multi-turn dialog between a human user and GPT-4V. 109 # to Published as a conference paper at ICLR 2024 How many baseballs are there? Let's count the baseballs in each bucket: 1. First bucket: 5 baseballs 2. Second bucket: 6 baseballs 3. Third bucket: 6 baseballs 4. Fourth bucket: 6 baseballs Total baseballs: 5 + 6 + 6 + 6 = 23 baseballs. Answer: 23. Your solution is wrong. You counted the baseballs in each bucket incorrectly. 2) | apologize for the oversight. Let's try counting again: 1. First bucket: 5 baseballs 2. Second bucket: 5 baseballs 3. Third bucket: 5 baseballs 4. Fourth bucket: 5 baseballs Total baseballs: 5 + 5 + 5 + 5 = 20 baseballs. Answer: 20. Figure 114: GPT-4V initially produces an incorrect solution due to errors in visually perceiving the number of baseballs. After receiving user feedback, it recalibrates and generates the accurate answer. 110 | 2310.02255#246 | 2310.02255#248 | 2310.02255 | [
"2302.13971"
] |
2310.02255#248 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Published as a conference paper at ICLR 2024 â eee Tew 13057 91 18131581419? (BiBiaiele| â â SSâ â â â Which number comes next? The pattern seems to involve prime numbers. Let's break it down: S 1 and 3 are the first and second prime numbers. 5 and 7 are the third and fourth prime numbers. 9 (not a prime) but followed by 11 which is the fifth prime number. 13 is the sixth prime number and 15 (not a prime) followed by 17 which is the seventh prime number. 19 is the eighth prime number. Following this pattern, the next two prime numbers are 23 and 29. So, the next number in the sequence would be 2329. Your solution is incorrect. The pattern does not involve prime numbers. 2) | apologize for the oversight. Let's re-examine the pattern. S It seems the pattern is derived by concatenating consecutive numbers: 1 and 3 give 13. 5 and 7 give 57. 9 and 11 give 911. 13 and 15 give 1315. 17 and 19 give 1719. Following this pattern, the next numbers to be concatenated would be 21 and 23. So, the next number in the sequence would be 2123. Figure 115: GPT-4V correctly identifies the sequence of numbers in the images but struggles to discern the underlying pattern. With feedback from the user, GPT-4V re-evaluates the pattern and produces the correct response to the logical reasoning question. | 2310.02255#247 | 2310.02255#249 | 2310.02255 | [
"2302.13971"
] |
2310.02255#249 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 111 Published as a conference paper at ICLR 2024 Published as a conference paper at ICLR 2024 Figure 116: Initially, GPT-4V applies an incorrect theorem to address the scientific reasoning query. However, after receiving the appropriate domain knowledge hint from the user, it utilizes the correct theorem, combined with key visual cues from the images, to produce the right answer. 112 Published as a conference paper at ICLR 2024 Published as a conference paper at ICLR 2024 | 2310.02255#248 | 2310.02255#250 | 2310.02255 | [
"2302.13971"
] |
2310.02255#250 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Table 1: AgentBench evaluates 25 API-based or open-sourced LLMs on LLM-as-Agent challenges. Model Model Size Form Version Creator gpt-4 [56] undisclosed api 0613 3.5-turbo [55] undisclosed api 0613 OpenAl -davinci-003 [58] undisclosed api - text -davinci-002 [58] undisclosed api - claude [6] undisclosed api v1.3 claude-instant [6] undisclosed api vL1 text-bison-001 [3] undisclosed api - Google chatg1m2 [100; 21] undisclosed api . . . chatglm-6b [100; 21] 6B open . Tsinghua & Zhipu codegeex2-6b-chat [102] 6B open - vicuna-13b [14] 13B open . vicuna-7b [14] 7B open vi. LMSYS 1lama2-chat-13b [80] 13B open Llama2-chat-7b [80] 7B open openchat -8192-13b [85] 13B â open : openchat -13b [85] 13B open Tsinghua wizard1m-30b [93] 30B open a wizardcoder-15b [48] 15B open zt Microsoft wizardim-13b [93] 13B open koala-13b [26] 1BB open UC Berkeley internlm-chat-7b [79] 7B open InternLM Team baichuan-13b-chat [36] open baichuan-7b [36] 7B open dolly-12b [16] 12B open Databricks oasst-sft-4-12b [42] 12B open LAION Anthropic Meta Baichuan Inc. How many models in the table have a model size larger than 10B? From the table: | 2310.02255#249 | 2310.02255#251 | 2310.02255 | [
"2302.13971"
] |
2310.02255#251 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | Models with size larger than 10B: 1. chatglm2 [100; 21] 2. vicuna-13b [14] 3. vicuna-7b [14] 4. 11ama2-chat-13b [80] 5. openchat-8192-13b [85] 6. openchat-13b [85] 7. wizardlm-30b [93] 8. wizardim-13b [93] 9. koala-13b [26] 10. baichuan-13b-chat [36] Total: 10 models. Figure 117: Answering this question requires comprehending a complex table regarding AI models (such as identifying relevant rows and columns), and up-to-date knowledge of AI research, and the ability to perform advanced arithmetic reasoning. Without human in the loop, GPT-4V fails to produce the correct answer. | 2310.02255#250 | 2310.02255#252 | 2310.02255 | [
"2302.13971"
] |
2310.02255#252 | MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts | 113 Published as a conference paper at ICLR 2024 Published as a conference paper at ICLR 2024 Figure 118: (Continued) In this multi-turn human-AI conversation, the user acts as a tutor to help GPT-4V generate the correct answer . (1) Initially, the user directs GPT-4V to the â Model Sizeâ col- umn after observing its misdirection. Despite the guidance, GPT-4V still struggles. (2) Recognizing the modelâ s difficulty in identifying the â Model Sizeâ column, the user provides more detailed in- structions. This time, GPT-4V is on the right track but still makes some error in its response. (3) With further information from the user, GPT-4V corrects the remaining errors and generate the cor- rect answer. 114 Published as a conference paper at ICLR 2024 | 2310.02255#251 | 2310.02255#253 | 2310.02255 | [
"2302.13971"
] |
Subsets and Splits