"我可能没有表达清楚"：在推理时诊断大语言模型推理的动态不稳定性

摘要

大型语言模型（LLM）的推理失败通常仅在生成结束时进行测量，然而许多失败表现为过程层面的崩溃：模型在推理过程中"偏离主线"。我们研究是否能够通过标准API中可用的推理时观测值（词元对数概率），在无需任何训练或微调的情况下检测此类崩溃。我们定义了一个结合连续步骤分布偏移（JSD）和不确定性（熵）的简单不稳定性信号，通过峰值不稳定强度概括每个轨迹，并证明该信号能可靠预测失败。在GSM8K和HotpotQA数据集中，不稳定强度能以高于随机水平的AUC值预测错误答案，并随模型规模扩大呈现桶级准确率的单调下降。关键的是，我们发现不稳定性并非一概有害：早期不稳定性可能反映后续的稳定过程并得到正确答案（修正性不稳定），而晚期不稳定性更常导致最终失败（破坏性不稳定）——即使在峰值强度相近时也是如此。这表明可恢复性不仅取决于分布变化的强度，更取决于这种变化相对于剩余解码窗口的发生时机。该方法具有模型无关性、免训练性和可复现性，是作为诊断视角而非修正或控制机制提出的。

English

Reasoning failures in large language models (LLMs) are typically measured only at the end of a generation, yet many failures manifest as a process-level breakdown: the model "loses the thread" mid-reasoning. We study whether such breakdowns are detectable from inference-time observables available in standard APIs (token log probabilities), without any training or fine-tuning. We define a simple instability signal that combines consecutive-step distributional shift (JSD) and uncertainty (entropy), summarize each trace by its peak instability strength, and show that this signal reliably predicts failure. Across GSM8K and HotpotQA, instability strength predicts wrong answers with above-chance AUC and yields monotonic bucket-level accuracy decline at scale across model sizes. Crucially, we show that instability is not uniformly harmful: early instability can reflect subsequent stabilization and a correct final answer (corrective instability), whereas late instability is more often followed by failure (destructive instability), even at comparable peak magnitudes, indicating that recoverability depends not only on how strongly the distribution changes but also on when such changes occur relative to the remaining decoding horizon. The method is model-agnostic, training-free, and reproducible, and is presented as a diagnostic lens rather than a corrective or control mechanism.

"我可能没有表达清楚"：在推理时诊断大语言模型推理的动态不稳定性

"I May Not Have Articulated Myself Clearly": Diagnosing Dynamic Instability in LLM Reasoning at Inference Time

摘要

Support