An Interpretable Memory Decision Controller for LLM Agents Based on Three-Signal Complementarity: Decoupling Confidence and Consistency
An interpretable memory decision controller reduces hallucination rate by 56.04% using three-signal complementarity.
Key Findings
Methodology
The paper introduces the Memory Decision Layer (MDL), which uses a three-signal complementary encoder to integrate relevance, reliability, and task risk signals. It employs QR-based orthogonal subspace projection and a meta-working-memory signal to create an interpretable decision representation quantifying the trustworthiness of retrieved memories. MDL explicitly decouples confidence from consistency and introduces risk inversion and explicit abstention.
Key Results
- MDL reduces hallucination rate from 53.0% to 23.3% under conflicting memories, approaching zero in high-risk scenarios.
- Validated on TruthfulQA and HaluEval datasets, MDL performs excellently across mainstream LLMs.
- Compared to standard RAG, MDL triggers abstention in high-risk scenarios, eliminating hallucinations.
Significance
This research is significant for both academia and industry, addressing the trust decision problem in memory systems of large language models during multi-turn interactions and complex task solving, especially under memory conflicts.
Technical Contribution
MDL achieves zero-parameter deployment through geometric operations, significantly improving decision speed and safety, offering new engineering possibilities and theoretical guarantees distinct from existing SOTA methods.
Novelty
This is the first to insert a trust decision layer between memory retrieval and generation, explicitly decoupling confidence from consistency, combining risk inversion with explicit abstention, filling a gap in existing research.
Limitations
- MDL's performance in low-risk scenarios needs further optimization, as it may be overly conservative in some cases.
- The task risk assessment relies on predefined risk coefficients, which may lack flexibility.
Future Work
Future work could explore dynamic risk assessment mechanisms and applications across more datasets and task scenarios to enhance MDL's generalization and adaptability.
AI Executive Summary
Large language models (LLMs) rely on external memory systems for multi-turn interactions, but existing retrieval-augmented generation (RAG) methods often lead to hallucination issues under memory conflicts. This paper introduces a new Memory Decision Layer (MDL) that integrates relevance, reliability, and task risk signals using a three-signal complementary encoder to create an interpretable decision representation quantifying the trustworthiness of retrieved memories.
The core innovation of MDL is the explicit decoupling of confidence from consistency, along with the introduction of risk inversion and explicit abstention mechanisms. Experimental results show that MDL reduces hallucination rates by 56.04% under conflicting memories and approaches zero in high-risk scenarios. This method requires no trained parameters and relies solely on geometric operations, achieving decision speeds 50 times faster than the embedding retrieval step.
This research is significant for both academia and industry, addressing the trust decision problem in memory systems of large language models during multi-turn interactions and complex task solving, especially under memory conflicts. Future work could explore dynamic risk assessment mechanisms and applications across more datasets and task scenarios to enhance MDL's generalization and adaptability.
Deep Analysis
Background
Large language models (LLMs) rely on external memory systems for multi-turn interactions and complex task solving. Existing retrieval-augmented generation (RAG) methods focus on efficient retrieval but pay little attention to the trust decision of retrieved memories. Under memory conflicts, RAG methods often lead to hallucination issues, affecting the reliability of model outputs.
Core Problem
Existing RAG methods often lead to hallucination issues under memory conflicts, affecting the reliability of model outputs. How to make trust decisions between retrieval and generation to reduce hallucination issues is a pressing problem.
Innovation
This paper introduces a new Memory Decision Layer (MDL) that integrates relevance, reliability, and task risk signals using a three-signal complementary encoder to create an interpretable decision representation quantifying the trustworthiness of retrieved memories. MDL explicitly decouples confidence from consistency and introduces risk inversion and explicit abstention mechanisms.
Methodology
- �� Three-signal complementary encoder: integrates relevance, reliability, and task risk signals.
- �� QR-based orthogonal subspace projection: creates interpretable decision representation.
- �� Explicit decoupling of confidence and consistency: enhances auditability and safety.
- �� Risk inversion and explicit abstention: triggers abstention in high-risk scenarios.
Experiments
Experiments were conducted on TruthfulQA and HaluEval datasets to validate MDL's performance across mainstream LLMs. Baselines include no memory, standard RAG, and other learning methods. Key metrics include hallucination rate and decision speed.
Results
MDL reduces hallucination rate from 53.0% to 23.3% under conflicting memories, approaching zero in high-risk scenarios. Compared to standard RAG, MDL triggers abstention in high-risk scenarios, eliminating hallucinations.
Applications
MDL can be applied in scenarios requiring high reliability and safety, such as healthcare, legal, and financial fields, helping reduce hallucination issues and improve the trustworthiness of model outputs.
Limitations & Outlook
MDL's performance in low-risk scenarios needs further optimization, as it may be overly conservative in some cases. The task risk assessment relies on predefined risk coefficients, which may lack flexibility. Future work could explore dynamic risk assessment mechanisms.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen, and the memory system is like your pantry. You choose the most relevant ingredients, but sometimes these ingredients might be expired or unsuitable for the current recipe. MDL acts like a smart assistant, helping you decide which ingredients are fresh and reliable, and alerting you to those that might ruin the dish. This way, you can create tastier dishes without failing due to wrong ingredients.
ELI14 Explained like you're 14
Imagine you're playing a game, and you have a backpack full of items. Every time you want to use an item, MDL acts like a smart assistant, helping you pick the best item for the current task. If an item might make you lose the game, it alerts you to avoid using it. This way, you can complete tasks better without failing due to wrong items.
Glossary
Memory Decision Layer
A controller situated between retrieval and generation to assess the trustworthiness of retrieved memories.
Used in this paper to reduce hallucination issues.
Three-Signal Complementary Encoder
An encoder that integrates relevance, reliability, and task risk signals.
Used to create interpretable decision representation.
QR-Based Orthogonal Subspace Projection
A geometric operation used to create interpretable decision representation.
Used in MDL for signal integration.
Hallucination
Model outputs that are inconsistent with facts.
Occurs under memory conflicts.
Risk Inversion
A mechanism that reduces hallucination by lowering the activation of high-risk signals.
Triggers abstention in high-risk scenarios.
Open Questions Unanswered questions from this research
- 1 How to optimize MDL's performance in low-risk scenarios to avoid overly conservative decisions?
- 2 How to dynamically assess task risk to improve MDL's flexibility and adaptability?
Applications
Immediate Applications
Medical Decision-Making
Apply MDL in medical scenarios to reduce misdiagnosis and erroneous advice, improving diagnostic accuracy.
Long-term Vision
Intelligent Legal Advisor
Apply MDL in the legal field to provide more reliable legal advice, reducing the risk of misjudgment.
Abstract
Memory systems for large language models have focused predominantly on efficient retrieval, whereas the decision of whether retrieved memories should be trusted has received comparatively little attention. When the memory store contains conflicting positions, standard retrieval-augmented generation (RAG) blindly injects memories and amplifies hallucinations: in models susceptible to memory injection, the RAG hallucination rate under conflicting memories is markedly higher than that of a memory-free baseline. Inspired by memory signaling mechanisms in the prefrontal cortex, we propose the Memory Decision Layer (MDL), a zero-parameter memory decision controller situated between the retrieval and generation stages. Its core is a three-signal complementary encoder that fuses relevance, reliability, and task risk through QR-based orthogonal subspace projection and a meta-working-memory signal into an interpretable decision representation that quantifies the trustworthiness of retrieved memories. Building on this encoder, MDL explicitly decouples confidence from consistency and introduces risk inversion and explicit abstention. Evaluations on mainstream large language models and multiple open-source datasets show that MDL reduces the hallucination rate under conflicting memories by about 56.04% in general scenarios and approaches zero hallucination in high-risk scenarios. The controller is fully white-box: it relies purely on geometric operations, requires no trained parameters, and adds only about 0.14 ms per decision -- roughly 50x faster than the embedding-retrieval step that precedes it and four to five orders of magnitude faster than an LLM self-evaluation call.