An Interpretable Memory Decision Controller for LLM Agents Based on Three-Signal Complementarity: Decoupling Confidence and Consistency

Imagine you're building a large language model that needs to pull in information from a memory store to generate responses. The challenge arises when this memory contains conflicting information; the model might blindly use these memories, leading to hallucinations or incorrect outputs. This issue is particularly pronounced in retrieval-augmented generation (RAG) systems, where the model's reliance on potentially unreliable memories can amplify errors, creating a situation where the model's outputs are less trustworthy than if it had no memory at all. This phenomenon is known as hallucination, where the model generates false or misleading information based on the conflicting memories it retrieves. To address this, the authors propose a Memory Decision Layer (MDL), which acts as a filter between memory retrieval and response generation. The MDL evaluates the trustworthiness of retrieved memories by considering their relevance, reliability, and the risk associated with using them. It employs a geometric approach to create a decision representation that helps the model determine when to trust a memory and when to abstain from using it altogether. This method not only reduces the hallucination rate significantly—by over 56% in general scenarios—but also approaches zero in high-risk situations. Compared to previous methods, the MDL is efficient, requiring no additional training and operating much faster than traditional evaluation methods, making it a practical solution for builders looking to enhance the reliability of language models.

The introduction of a zero-parameter memory decision controller is a significant advancement in managing memory trustworthiness.

The evaluations demonstrate a substantial reduction in hallucination rates across various scenarios, supporting the method's effectiveness.