aiqbench analysis
This paper tackles a practical cost problem: long chain-of-thought reasoning makes LLM inference expensive. Its method trains smaller models to write compact "draft" reasoning using supervised fine-tuning and two-stage reinforcement learning, without post-hoc compression or a length penalty. The reported result, an 82.6% shorter response on MATH500 for a 3.4-point accuracy drop, comes from a small training setup, so readers should check how it holds up across models and benchmarks before assuming it generalizes.
Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs
arXiv:2603.00578v2 Announce Type: replace-cross Abstract: Long chain-of-thought~(CoT) reasoning improves the problem-solving capability of large reasoning models~(LRMs), but often incurs substantial inference cost and overthinking. Existing efficiency methods primarily rely on prompting, post-hoc compression, or length-regularized optimization; these approaches can be sensitive to model scale, depend on external estimators, or under-allocate reasoning to difficult instances. We propose \textbf{Draft-Thinking}, a lightweight adaptation framework that learns a \textit{model-adapted draft CoT}: a compact, self-contained working derivation that directly supports the final answer. We identify a \textbf{Cross-Scale Draft Mismatch}: concise traces generated by a strong teacher can be too sparse for a smaller target model. To address this mismatch, \textbf{Progressive Draft Adaptation} first transfers a concise reasoning-form prior through supervised fine-tuning (SFT) and then applies two-stage reinforcement learning (RL) with progressively harder data and relaxed response budgets to calibrate the reasoning granularity to the target model. The method requires neither post-hoc trajectory compression nor an explicit length-penalty reward. At inference time, draft, adaptive, and step-by-step prompts provide efficiency-oriented, balanced, and detailed reasoning modes, respectively. The step-by-step mode largely preserves the original model's accuracy while using fewer tokens. With only 847 training problems and 120 RL optimization steps, Draft-Thinking substantially reduces reasoning cost across six mathematical and scientific reasoning benchmarks. On MATH500, its draft mode reduces the average response length from 5,668 to 986 tokens, an 82.6\% reduction, while achieving 90.6\% accuracy compared with 93.0\% for the original long CoT model.
Abstract:Long chain-of-thought~(CoT) reasoning improves the problem-solving capability of large reasoning models~(LRMs), but often incurs substantial inference cost and overthinking. Existing efficiency methods primarily rely on prompting, post-hoc compression, or length-regularized optimization; these approaches can be sensitive to model scale, depend on external estimators, or under-allocate reasoning to difficult instances. We propose \textbf{Draft-Thinking}, a lightweight adaptation framework that learns a \textit{model-adapted draft CoT}: a compact, self-contained working derivation that directly supports the final answer. We identify a \textbf{Cross-Scale Draft Mismatch}: concise traces generated by a strong teacher can be too sparse for a smaller target model. To address this mismatch, \textbf{Progressive Draft Adaptation} first transfers a concise reasoning-form prior through supervised fine-tuning (SFT) and then applies two-stage reinforcement learning (RL) with progressively harder data and relaxed response budgets to calibrate the reasoning granularity to the target model. The method requires neither post-hoc trajectory compression nor an explicit length-penalty reward. At inference time, draft, adaptive, and step-by-step prompts provide efficiency-oriented, balanced, and detailed reasoning modes, respectively. The step-by-step mode largely preserves the original model's accuracy while using fewer tokens. With only 847 training problems and 120 RL optimization steps, Draft-Thinking substantially reduces reasoning cost across six mathematical and scientific reasoning benchmarks. On MATH500, its draft mode reduces the average response length from 5,668 to 986 tokens, an 82.6\% reduction, while achieving 90.6\% accuracy compared with 93.0\% for the original long CoT model.
Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2603.00578 [cs.AI]
(or arXiv:2603.00578v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2603.00578 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 326 words
Related coverage
Why I Left Google DeepMind
LessWrong · Jul 15, 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · Sep 17, 2026
Game Design Creativity evals
LessWrong · Oct 12, 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · Sep 18, 2026