发现长思维链中错误固化现象，提出自适应重启机制提升推理正确率。
Thinking Traps in Long Chain-of-Thought: A Measurable Study and Trap-Aware Adaptive Restart
- 通过轨迹分析识别出错误固化陷阱，定位需截断的思维节点。
- 在DAPO-MATH上89%的失败案例存在此类陷阱，改进后推理准确率显著提升。
- 无需微调模型参数，适合高难度数学与科学推理任务使用。
通过延长测试时计算量的长思维链（Long-CoT）能显著增强推理能力，但过长生成未必带来正确性：一旦早期犯错，模型可能持续构建一个自洽却错误的前缀。我们通过细粒度轨迹分析，发现了‘思维陷阱’——一种由前缀主导的死锁状态，后续反思、尝试或验证均无法修正根本错误。在精心筛选的DAPO-MATH子集上，89%的失败案例表现出此类陷阱。为此，我们提出TAAR（Trap-Aware Adaptive Restart）测试时控制框架，训练诊断策略从部分轨迹中预测两个信号：陷阱位置索引与逃逸概率。推理时，TAAR在预判陷阱前截断轨迹，并自适应重启解码；对严重困局，采用更强扰动，包括更高温度重采样及可选结构化重启后缀。在挑战性数学与科学推理基准（AIME24、AIME25、GPQA-Diamond、HMMT25、BRUMO25）上的实验表明，TAAR在不微调基础模型参数的情况下有效提升推理性能。
原文摘要 · Abstract (English)
Scaling test-time compute via Long Chain-of-Thought (Long-CoT) significantly enhances reasoning capabilities, yet extended generation does not guarantee correctness: after an early wrong commitment, models may keep elaborating a self-consistent but incorrect prefix. Through fine-grained trajectory analysis, we identify Thinking Traps, prefix-dominant deadlocks where later reflection, alternative attempts, or verification fails to revise the root error. On a curated subset of DAPO-MATH, 89\% of failures exhibit such traps. To solve this problem, we introduce TAAR (Trap-Aware Adaptive Restart), a test-time control framework that trains a diagnostic policy to predict two signals from partial trajectories: a trap index for where to truncate and an escape probability for whether and how strongly to intervene. At inference time, TAAR truncates the trajectory before the predicted trap segment and adaptively restarts decoding; for severely trapped cases, it applies stronger perturbations, including higher-temperature resampling and an optional structured reboot suffix. Experiments on challenging mathematical and scientific reasoning benchmarks (AIME24, AIME25, GPQA-Diamond, HMMT25, BRUMO25) show that TAAR improves reasoning performance without fine-tuning base model parameters.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。