大型推理模型何时能“节省思考”？推理中行为分歧的机制分析
When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in Reasoning
AI总结：
本研究分析大型推理模型的“过度思考”行为，提出注意力干预方法调控该行为，揭示指令遵循、推理效率与正确性间的权衡。
摘要：
大型推理模型（LRMs）在复杂任务上取得了显著成功，但其“过度思考”的倾向导致效率低下。尽管“save-thinking”提示旨在缓解该问题，我们发现LRMs仍频繁进入“Still-thinking”模式，而非预期的“No-thinking”模式，尤其在困难查询上。为分析该行为分歧，我们从三个视角研究LRMs：思考终止边界的置信度、内部注意力分布的分歧、以及提示片段间的注意力分配。我们发现高困惑度与更晚的Still-thinking行为相关，且Still-thinking案例会对原始问题分配更多注意力。基于这些观察，我们提出一种注意力干预方法来调控该行为。该干预虽抑制了显式思考，但也导致准确率下降，表明被抑制的推理行为通常对正确性有用。本工作为该行为提供了置信度和注意力层面的证据，凸显了指令遵循、推理效率与推理正确性之间的权衡。
英文摘要：
Large reasoning models (LRMs) have achieved remarkable success on complex tasks, yet their tendency to "overthink" leads to inefficiencies. Although "save-thinking" prompts are intended to mitigate this issue, we find that LRMs still frequently enter the "Still-thinking" mode instead of the expected "No-thinking" mode, especially on difficult queries. To analyze this behavioral divergence, we examine LRMs from three perspectives: confidence at the thinking-termination boundary, divergence in internal attention distributions, and attention allocation across prompt segments. We find that high perplexity is associated with later Still-thinking behavior, and that Still-thinking cases allocate more attention to the original question. Based on these observations, we propose an attention intervention method to regulate this behavior. While this intervention suppresses explicit thinking, it also causes a drop in accuracy, suggesting that the suppressed reasoning behavior is often useful for correctness. Our work provides confidence- and attention-level evidence for this behavior, highlighting the trade-off between instruction following, inference efficiency, and reasoning correctness.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
