让大模型根据题目难易自动决定思考长度，大幅提速且不丢精度。
DART: Difficulty-Adaptive Reasoning Truncation for Efficient Large Language Models
- 通过蒸馏强模型的简洁推理模式，构建可调节的思维风格连续体。
- 在GSM8K上实现81.2%的推理截断率，计算效率提升5.33倍。
- 无需强化学习，训练稳定，适合部署在资源受限场景的推理系统。
自适应推理对齐大语言模型（LLMs）的计算开销与问题内在难度至关重要。现有思维链方法虽提升推理能力，但盲目生成长解释，造成明显低效。现有强化学习方案存在训练不稳定、高度依赖奖励等问题。本文提出DART——一种监督式难度自适应推理截断框架，根据问题难度动态调整思考长度。通过从更强模型中蒸馏简洁推理模式，将其插值为连续的推理风格，并筛选兼顾正确性与紧凑性的最优训练数据，使模型学会何时‘停止思考’。在多个数学基准测试中，实验结果表明其兼具显著效率与高准确率：在DeepSeek-R1-Distill-Qwen-7B模型上，于GSM8K数据集实现81.2%的推理截断率，计算加速达5.33倍。DART提供了一种稳定通用的高效推理范式，推动大模型自适应智能的发展。
原文摘要 · Abstract (English)
Adaptive reasoning is essential for aligning the computational effort of large language models (LLMs) with the intrinsic difficulty of problems. Current chain-of-thought methods boost reasoning ability but indiscriminately generate long explanations, leading to evident inefficiency. However, existing reinforcement learning approaches to adaptive thinking remain unstable and heavily reward-dependent. Here we propose \textbf{DART}, a supervised \textbf{D}ifficulty-\textbf{A}daptive \textbf{R}easoning \textbf{T}runcation framework that adjusts thinking length according to problem difficulty. By distilling concise reasoning patterns from stronger models, interpolating them into a continuum of reasoning styles, and curating optimal training data that balances correctness and compactness, DART learns when to ``stop thinking''. Across multiple mathematical benchmarks, experimental results demonstrate its remarkable efficiency while preserving or improving accuracy, achieving a significant 81.2\% reasoning truncation (DeepSeek-R1-Distill-Qwen-7B on GSM8K dataset) with 5.33$\times$ computational acceleration. DART provides a stable and general paradigm for efficient reasoning, advancing the development of adaptive intelligence in LLMs.
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。