Investigating Tool-Memory Conflicts in Tool-Augmented LLMs
This addresses a critical reliability issue for developers and users of tool-augmented LLMs, though it is incremental as it builds on known knowledge conflict problems.
The paper identifies Tool-Memory Conflict (TMC) as a new type of knowledge conflict in tool-augmented LLMs where internal parametric knowledge contradicts external tool knowledge, particularly on STEM tasks, and finds that existing conflict resolution techniques fail to effectively resolve it.
Tool-augmented large language models (LLMs) have powered many applications. However, they are likely to suffer from knowledge conflict. In this paper, we propose a new type of knowledge conflict -- Tool-Memory Conflict (TMC), where the internal parametric knowledge contradicts with the external tool knowledge for tool-augmented LLMs. We find that existing LLMs, though powerful, suffer from TMC, especially on STEM-related tasks. We also uncover that under different conditions, tool knowledge and parametric knowledge may be prioritized differently. We then evaluate existing conflict resolving techniques, including prompting-based and RAG-based methods. Results show that none of these approaches can effectively resolve tool-memory conflicts.