Selective Memory Boosts LLM Scientific Reasoning

This study explores the nuanced role of conditional memory in large language models (LLMs) when performing scientific reasoning tasks. While external memory can provide crucial specialized knowledge, it can also introduce noise or interfere with the model's inherent reasoning capabilities. The researchers systematically analyzed when, where, and to what extent memory should be engaged to optimize performance. The core contribution is a "Knowledge Boundary-Aware Router," which uses task-specific input proxies to dynamically decide if memory should be activated, which internal model layers receive memory signals, and the strength of these signals. This selective approach aims to leverage beneficial memory contributions while mitigating potential regressions caused by inappropriate memory access. Experiments across various biological and chemical reasoning benchmarks, involving different LLM backbones and task types, confirm that memory effects are highly variable. The proposed router consistently outperforms static or random memory allocation, demonstrating that intelligent, selective memory management is crucial for reliable and accurate scientific reasoning in LLMs.