不训练模型，自动识别推理终点，大幅降低推理成本。
SyncThink: A Training-Free Strategy to Align Inference Termination with Reasoning Saturation
- 通过监测模型内部推理信号，动态决定何时停止思考。
- 在多个数据集上以656个词元实现62.00%准确率，速度提升3倍。
- 适合需要高效推理的长链条任务，尤其擅长防过度思考。
Chain-of-Thought（CoT）提示虽提升推理能力，但常生成冗长耗时的推理轨迹，显著增加推理开销。本文提出 SyncThink，一种无需训练、可即插即用的解码策略，可在不修改模型权重的前提下减少 CoT 开销。我们发现，答案词元对早期推理内容关注较弱，反而更依赖特殊标记“/think”，揭示存在信息瓶颈。基于此，SyncThink 监测模型自身的推理转换信号并主动终止推理。在 GSM8K、MMLU、GPQA 与 BBH 上，使用三个 DeepSeek-R1 深度蒸馏模型进行实验，SyncThink 在平均 Top-1 准确率达 62.00% 的同时，仅需生成 656 个词元和 28.68 秒延迟，相较完整 CoT 解码（61.22% 准确率，2141 词元，92.01 秒）效率大幅提升。在长周期任务如 GPQA 上，甚至可额外提升 8.1% 绝对准确率，有效防止过拟合式过度思考。
原文摘要 · Abstract (English)
Chain-of-Thought (CoT) prompting improves reasoning but often produces long and redundant traces that substantially increase inference cost. We present SyncThink, a training-free and plug-and-play decoding method that reduces CoT overhead without modifying model weights. We find that answer tokens attend weakly to early reasoning and instead focus on the special token "/think", indicating an information bottleneck. Building on this observation, SyncThink monitors the model's own reasoning-transition signal and terminates reasoning. Experiments on GSM8K, MMLU, GPQA, and BBH across three DeepSeek-R1 distilled models show that SyncThink achieves 62.00 percent average Top-1 accuracy using 656 generated tokens and 28.68 s latency, compared to 61.22 percent, 2141 tokens, and 92.01 s for full CoT decoding. On long-horizon tasks such as GPQA, SyncThink can further yield up to +8.1 absolute accuracy by preventing over-thinking.
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。