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

AI Digest

文章提出基于三信号互补性的可解释记忆决策控制器MDL，通过分离置信度与一致性显著降低LLM在冲突记忆下的幻觉率。

This paper introduces MDL, an interpretable memory decision controller that reduces hallucination rates in LLMs by 56.04% under conflicting memories through three-signal complementarity.

Key points

• MDL通过相关性、可靠性、任务风险三信号融合实现决策分离 MDL fuses relevance, reliability, and task risk through three-signal complementarity
• 白盒设计无需训练参数，决策速度达0.14ms Zero-parameter white-box design with 0.14ms decision latency
• 引入风险反转机制和显式回避策略 Introduces risk inversion and explicit abstention strategies
• 在主流LLM和多数据集验证效果显著 Validated across mainstream LLMs and multiple datasets
• 解决RAG方法在冲突记忆中的幻觉放大问题 Addresses hallucination amplification in conflicting memories

Takeaway: MDL通过创新决策机制显著提升LLM记忆处理的可靠性与安全性。 / MDL's innovative decision framework significantly enhances LLM memory processing reliability and safety.

Why it matters 为LLM记忆管理提供创新解决方案，可直接应用于提升模型可靠性与效率的场景。