Post
1870
š We're excited to introduce MemoryCode, a novel synthetic dataset designed to rigorously evaluate LLMs' ability to track and execute coding instructions across multiple sessions. MemoryCode simulates realistic workplace scenarios where a mentee (the LLM) receives coding instructions from a mentor amidst a stream of both relevant and irrelevant information.
š” But what makes MemoryCode unique?! The combination of the following:
ā Multi-Session Dialogue Histories: MemoryCode consists of chronological sequences of dialogues between a mentor and a mentee, mirroring real-world interactions between coworkers.
ā Interspersed Irrelevant Information: Critical instructions are deliberately interspersed with unrelated content, replicating the information overload common in office environments.
ā Instruction Updates: Coding rules and conventions can be updated multiple times throughout the dialogue history, requiring LLMs to track and apply the most recent information.
ā Prospective Memory: Unlike previous datasets that cue information retrieval, MemoryCode requires LLMs to spontaneously recall and apply relevant instructions without explicit prompts.
ā Practical Task Execution: LLMs are evaluated on their ability to use the retrieved information to perform practical coding tasks, bridging the gap between information recall and real-world application.
š Our Findings
1ļøā£ While even small models can handle isolated coding instructions, the performance of top-tier models like GPT-4o dramatically deteriorates when instructions are spread across multiple sessions.
2ļøā£ This performance drop isn't simply due to the length of the context. Our analysis indicates that LLMs struggle to reason compositionally over sequences of instructions and updates. They have difficulty keeping track of which instructions are current and how to apply them.
š Paper: From Tools to Teammates: Evaluating LLMs in Multi-Session Coding Interactions (2502.13791)
š¦ Code: https://github.com/for-ai/MemoryCode
š” But what makes MemoryCode unique?! The combination of the following:
ā Multi-Session Dialogue Histories: MemoryCode consists of chronological sequences of dialogues between a mentor and a mentee, mirroring real-world interactions between coworkers.
ā Interspersed Irrelevant Information: Critical instructions are deliberately interspersed with unrelated content, replicating the information overload common in office environments.
ā Instruction Updates: Coding rules and conventions can be updated multiple times throughout the dialogue history, requiring LLMs to track and apply the most recent information.
ā Prospective Memory: Unlike previous datasets that cue information retrieval, MemoryCode requires LLMs to spontaneously recall and apply relevant instructions without explicit prompts.
ā Practical Task Execution: LLMs are evaluated on their ability to use the retrieved information to perform practical coding tasks, bridging the gap between information recall and real-world application.
š Our Findings
1ļøā£ While even small models can handle isolated coding instructions, the performance of top-tier models like GPT-4o dramatically deteriorates when instructions are spread across multiple sessions.
2ļøā£ This performance drop isn't simply due to the length of the context. Our analysis indicates that LLMs struggle to reason compositionally over sequences of instructions and updates. They have difficulty keeping track of which instructions are current and how to apply them.
š Paper: From Tools to Teammates: Evaluating LLMs in Multi-Session Coding Interactions (2502.13791)
š¦ Code: https://github.com/for-ai/MemoryCode
