推测性思考：在推理时利用大模型指导增强小模型推理能力
Speculative Thinking: Enhancing Small-Model Reasoning with Large Model Guidance at Inference Time
AI总结：
提出一种无需训练的推测性思考框架，通过让大推理模型在推理层面引导小模型，在提升小模型推理准确率的同时缩短输出长度。
摘要：
近期进展利用后训练来增强模型推理性能，这通常需要昂贵的训练流程，并且仍然存在低效、输出过长的问题。我们提出推测性思考，一种无需训练的框架，使大推理模型在推理层面引导小模型进行推理，区别于在词元层面操作的推测解码。我们的方法基于两个观察：（1）支持推理的词元（如“wait”）经常出现在结构分隔符（如“\n\n”）之后，作为反思或继续的信号；（2）大模型对反思行为有更强的控制，减少不必要的回溯同时提高推理质量。通过策略性地将反思步骤委托给能力更强的模型，我们的方法显著提升了推理模型的推理准确率，同时缩短了输出。在32B推理模型的辅助下，1.5B模型在MATH500上的准确率从83.2%提升至89.4%，实现了6.2%的大幅提升。同时，平均输出长度从5439个词元减少到4583个词元，下降了15.7%。此外，当应用于非推理模型（Qwen-2.5-7B-Instruct）时，我们的框架在相同基准上将准确率从74.0%提升至81.8%，实现了7.8%的相对提升。
英文摘要：
Recent advances leverage post-training to enhance model reasoning performance, which typically requires costly training pipelines and still suffers from inefficient, overly lengthy outputs. We introduce Speculative Thinking, a training-free framework that enables large reasoning models to guide smaller ones during inference at the reasoning level, distinct from speculative decoding, which operates at the token level. Our approach is based on two observations: (1) reasoning-supportive tokens such as "wait" frequently appear after structural delimiters like "\n\n", serving as signals for reflection or continuation; and (2) larger models exhibit stronger control over reflective behavior, reducing unnecessary backtracking while improving reasoning quality. By strategically delegating reflective steps to a more capable model, our method significantly boosts the reasoning accuracy of reasoning models while shortening their output. With the assistance of the 32B reasoning model, the 1.5B model's accuracy on MATH500 increases from 83.2% to 89.4%, marking a substantial improvement of 6.2%. Simultaneously, the average output length is reduced from 5439 tokens to 4583 tokens, representing a 15.7% decrease. Moreover, when applied to a non-reasoning model (Qwen-2.5-7B-Instruct), our framework boosts its accuracy from 74.0% to 81.8% on the same benchmark, achieving a relative improvement of 7.8%.
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。
