衡量和治愈推理刚性：从装饰性推理链到真正的忠实
Measuring and curing reasoning rigidity: from decorative chain-of-thought to genuine faithfulness
AI总结：
本文提出SLRC指标和LC-CoSR方法，通过Lyapunov稳定性保证减少推理刚性，发现高SLRC模型易受阿谀影响，并提出RIS评分预测错误检测。
摘要：
语言模型越来越多地通过在回答前写出逐步推理来展示其能力。但这些步骤是否真正被使用，还是答案在推理开始前就固定了？我们引入了步骤级推理能力（SLRC）指标，并证明其是一致的因果估计器（定理1）。我们提出LC-CoSR训练方法，通过Lyapunov稳定性保证直接减少刚性。评估16个前沿模型（o4-mini，GPT-5.4，Claude Opus，Grok-4，DeepSeek-R1，Gemini 2.5 Pro等）在六个领域（N=133-500）时，发现推理分为三种模式。OpenAI的o4-mini在五个任务中显示74-88%的步骤必要性（73.8-88.3%），是本研究中SLRC最高的。关键区别在于基于强化学习的推理训练，而非思考令牌：Grok-4的推理模式比非推理模式的忠实度更低（1.4% vs 7.2%必要性）。我们发现一个忠实度悖论——高SLRC模型更易受阿谀影响，并提出忠实度评分（RIS = SLRC x (1-Sycophancy)），显著预测错误检测（rho=0.66，p=0.026）。LC-CoSR在无外部模型依赖的情况下，实现比FARL和CSR基线少2.6倍的负奖励。
英文摘要：
Language models increasingly show their work by writing step-by-step reasoning before answering. But are these steps genuinely used, or is the answer rigid - fixed before reasoning begins? We introduce the Step-Level Reasoning Capacity (SLRC) metric and prove it is a consistent causal estimator (Theorem 1). We propose LC-CoSR, a training method with Lyapunov stability guarantees that directly reduces rigidity. Evaluating 16 frontier models (o4-mini, GPT-5.4, Claude Opus, Grok-4, DeepSeek-R1, Gemini 2.5 Pro, and others) across six domains at N=133-500, we find reasoning falls into three modes. OpenAI's o4-mini shows 74-88% step necessity on five of six tasks (73.8-88.3%) - the highest SLRC in our study. The critical differentiator is RL-based reasoning training, not thinking tokens: Grok-4's reasoning mode shows lower faithfulness than its non-reasoning mode (1.4% vs 7.2% necessity). We discover a faithfulness paradox - high-SLRC models are more susceptible to sycophancy - and propose the Reasoning Integrity Score (RIS = SLRC x (1-Sycophancy)), which significantly predicts error detection (rho=0.66, p=0.026). LC-CoSR achieves 2.6x less negative reward than FARL and CSR baselines without external model dependencies.
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。
