cs.CL提交于 2026-10-21已译

多步归一：LLM 智能体推理收缩与安全记忆复用的极限

Many Steps into One: Reasoning Contraction and the Limits of Safe Memory Reuse for LLM Agents

Jianghong Zhou · Hao Rong · Chen Lin · Daoan Zhang

中文摘要

思维模板、技能库及智能体记忆通过推理收缩（reasoning contraction）将多步推理简化为单步。收缩记忆在事实变更后会静默输出错误答案，而基于成功/失败计数的系统难以有效检测此类变更。理论证明表明，对称计数器需达到误报时域的量级才能察觉部分变更，未截断计数器甚至可能无法察觉。在五个 LLM 处理多跳事实编辑任务的实验中，ExpeL 和 ACE 风格计数器在相同信息下，陈旧答案数量比认证记忆多 7.0-13.5 倍；在带精确验证器的代码任务中多 35-73 倍。安全性存在代价，认证负载（certification load）决定了安全记忆可跳过的推理量，并在 Zipf 需求下形成缩放定律。该方案通过似然比加权失败、对实时推理进行效用认证、事实变更时重启证书，并以似然比检测器作为未公告变更的安全网，大编辑时则直接失效记忆。

关键要点

01指出基于成功/失败计数的记忆系统无法有效检测事实变更，导致智能体静默输出错误答案。
02理论证明对称计数器需达到误报时域的量级才能察觉部分变更，未截断计数器可能永远无法察觉。
03实验显示 ExpeL 和 ACE 风格计数器在多跳及代码任务中，陈旧答案数量比认证记忆多 7.0-73 倍。
04提出基于似然比加权的认证机制，通过重启证书和失效策略应对事实变更，平衡安全性与推理效率。

解读

尚无解读。

原始英文摘要

arXiv:2610.17544v2 Announce Type: replace Abstract: Thought templates, skill libraries, agent memories and caches turn many reasoning steps into one. We call this reasoning contraction. A contracted memory saves reasoning steps and errors until its facts change, then quietly answers wrong. Most systems trust memories by counting successes and failures. We prove that such counters are badly tuned change detectors: a symmetric counter notices a partial change only after a power of its false-alarm horizon, and an unclipped one may never notice it. On five LLMs answering multi-hop questions over edited facts, ExpeL- and ACE-style counters serve 7.0-13.5x more stale answers than certified memories given the same information, and on code tasks with an exact verifier 35-73x more. Safety also has a price. When facts change and each memory is judged only by its own verified outcomes, one number, the certification load, sets how much reasoning any safe memory can skip, and under Zipf demand it yields a scaling law. Our recipe weighs each memory's failures by its own likelihood ratio, certifies its utility against live reasoning, and restarts the certificate when facts change, with the likelihood-ratio detector as the safety net for unannounced changes; for large edits it also invalidates the memory.