arXiv：cs.CL· Danyang Zhao, Sendong Zhao, Yanrui Du, Haochun Wang, Ming Ma, Wenhao Li, Jingbo Wang, Jiatong Liu, Bing Qin·· 10 小时前
PULSE：面向医学推理的步级检索时机控制方法
Taking the PULSE of Medical Reasoning: Step-Level Retrieval Timing for Knowledge-Grounded Inference
AI 导读
PULSE 是一种轻量级探测方法，利用模型内部表示作为内在信号，实现推理步级的检索触发控制，判断每步该依赖内部知识还是调用外部证据。该方法基于错误步骤的层间波动显著高于正确步骤这一发现，训练紧凑探针检测事实错误并触发检索与步骤重生成。在四个医学基准上，PULSE 定时检索精度与逐步穷举检索相当，检索调用减少 72%，并可迁移至未见过的法律领域。
正文
Abstract:Reliable medical reasoning requires LLMs to dynamically access external biomedical evidence to improve factual trustworthiness and reasoning accuracy. However, indiscriminate retrieval can introduce noise and latency, raising a critical retrieval-timing problem: at each reasoning step, the system must decide whether the model's internal knowledge is reliable or external evidence is needed. We propose PUlSE, a lightweight probing method that uses the model's internal representations as intrinsic signals for step-level retrieval control. Our posterior trajectory analysis reveals that erroneous steps exhibit substantially higher inter-layer volatility than correct ones, suggesting unstable internal knowledge. Based on this finding, PUlSE trains a compact probe on inter-layer geometric features to detect factual errors upon step completion, and uses it as an adaptive trigger for retrieval and step regeneration. Across four medical benchmarks, PUlSE-timed retrieval matches exhaustive per-step retrieval in accuracy while reducing retrieval calls by 72%, and can transfer to an unseen legal domain, indicating potential domain-agnostic generality. Overall, our study provides an initial step toward shifting retrieval-trigger design from surface-level confidence estimates to internal activation dynamics for adaptive access.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.20292 [cs.CL] |
| (or arXiv:2610.20292v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.20292 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Danyang Zhao [view email]
[v1]
Wed, 14 Oct 2026 14:47:03 UTC (6,462 KB)
来源：arXiv：cs.CL · arxiv.org