Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs

arXiv:2603.00578 · cs.AI, cs.CL · Submitted 2026-02-28 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs".

Jane: The gist The Draft-Thinking method guides large reasoning models to first learn a concise draft-style reasoning structure that retains only the critical reasoning steps,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper, "Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs," and it looks like they’re tackling a big problem with long chain-of-thought prompting.

Jane: Right, so the main thing is that existing chain-of-thought methods tend to cause models to overthink, which means they use up way too much reasoning budget for a given question.

Lu: Exactly. The authors propose Draft-Thinking as a way to guide models to first learn this concise draft structure that only keeps the critical reasoning steps, and then they try to get that pattern stable through some sort of curriculum learning process.

Meng: So it’s not just about making them shorter; it’s about teaching them *what* reasoning steps are actually important so they don't waste tokens on stuff that doesn't matter for the final answer.

Lalam: I see. It sounds like they are trying to make the selection of high-value steps an endogenous capability of the large reasoning models themselves, rather than just relying on a fixed prompt structure or compression technique.

Tom: That’s what Page one says, they propose this method to guide models to first learn a concise draft-style reasoning structure that retains only the critical reasoning steps > <ref:2603.00578#pg1,models to first learn a concise draft-style reasoning structure that retains>

Jane: And then they use progressive curriculum learning to stably internalize this efficient pattern as the model's capability scales >

Lu: The authors also mention that Draft-Thinking introduces adaptive prompting, which elevates reasoning depth to a flexible, model-selectable behavior >

Meng: So instead of forcing one fixed way of thinking, the model can switch between draft reasoning mode for simpler questions and long chain-of-thought mode for more challenging ones >

Tom: That flexibility is key. It means they aren't binding models to uniformly high-budget trajectories when they can decide how deep to go based on the problem complexity >

Jane: And they show this in their results on the Math500 dataset, where Draft-Thinking achieves comparable or better accuracy with a smaller token budget compared to methods like ThinkPrune or CTS >

Lu: They are showing that Draft-Thinking can achieve comparable accuracy while using fewer tokens, and they treat efficient reasoning as something learnable and a decidable budget instead of just binding the models to uniformly high-budget trajectories >

Title and authors: Tom: I'm interested in how they do that learning part. It sounds like it’s not just a single training step, but this whole curriculum learning thing >

Jane: Right, and the idea is that they build this efficient reasoning framework first through supervised fine-tuning or SFT, and then use black-box distillation to extract those implicit structured inference patterns from stronger models >

Meng: Distillation sounds like a way to get those good reasoning patterns from a bigger model without needing all the data for that stronger model >

Lalam: That process leads into what they call Progressive Curriculum Learning, where they first fit draft-style reasoning structures at the behavioral level through SFT and then leverage distillation >

Tom: It’s about creating this reliable low-budget reasoning framework before introducing a two-stage reinforcement learning process to gradually expand the problem-solving capability >

Jane: That expansion stage is interesting because it’s designed to stably learn hard tasks that need deep reasoning while still keeping that preference for low budget efficiency >

Lu: And they do this by using an Incremental Length Expansion Strategy, which progressively relaxes the maximum generation length Lmax across multiple stages of reinforcement learning >

Tom: But they mention a specific detail about why they do this; imposing a large Lmax at the initial RL stage causes the model to revert to its original long chain-of-thought behavior even under the draft prompt >

Jane: That makes sense. It shows that if you force it to be long right away, it forgets what Draft-Thinking taught it about being concise >

Meng: So by expanding Lmax in stages, they can achieve progressive improvement while preserving that concise reasoning form, using two stages with Lmax of three thousand and then six thousand for the AIME24 dataset > <ref:2603.00578#pg2>

Lalam: That means they are systematically training the model to internalize those essential structures through optimization rather than just relying on predefined compression ratios or length penalties >

Tom: And when it comes to using it, at inference time, you can choose between Draft mode which maximizes token efficiency and long CoT mode which maximizes accuracy >

Title and authors: Jane: Or they have this hybrid instance adaptive prompt that lets the model assess the question's difficulty and adaptively choose between both modes >

Lu: That adaptive prompting is what allows them to balance token efficiency with high accuracy by leveraging the strengths of each approach depending on what’s needed for a specific question >

Meng: I wonder about the cost of running this system. Training it through SFT and then through distillation and staged RL sounds like it requires a lot of setup work >

Tom: The paper shows that Draft-Thinking achieves significantly superior token efficiency across all mathematical reasoning benchmarks, reaching an EFF of nine point one eight on MATH500 for the 32B model >

Jane: That’s a big number compared to other RL methods they compared it against, like ThinkPrune and DR.SAF, and even offline methods on larger models >

Lu: The authors also note that Draft-Thinking simultaneously enhances the model’s long CoT accuracy while substantially boosting its token efficiency after three stages of Draft-Thinking training >

Meng: So if I were deploying this, I'd be looking at those efficiency gains and making sure it doesn't break at scale when handling really complex problems >

Lalam: The authors also mention a limitation: on the most difficult AIME2025 dataset, the accuracy gap between draft and long-CoT modes is larger than on other sets >

Tom: They suspect that complex problems require a larger maximum sequence length during draft reasoning training, which suggests they aren't perfect yet >

Jane: So to wrap up Draft-Thinking: it successfully internalizes efficient reasoning as an endogenous capability, allows for flexible scheduling of the reasoning budget based on task complexity, and uses progressive curriculum learning to stably achieve that co-evolution of efficient structure and deep capability >

Lu: The main contribution is treating efficient reasoning as a learnable form rather than binding models to uniformly high-budget trajectories >

Meng: It’s a solid framework for getting better performance with less computational cost, provided the training pipeline works reliably across different problem sets >

Lalam: We should remember this paper, Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs, because it shows how to move toward more efficient reasoning structures in large models >

The paper's summary: Tom: So, to recap, Draft-Thinking is this method that teaches large reasoning models to first learn a short draft structure that just keeps the critical steps, and then makes them internalize that way through a learning process.

Jane: Right, so it’s about making the model decide which thinking steps are actually important so it doesn't waste all its processing power on fluff when solving a problem.

Lu: It frames reasoning efficiency as something the model learns by constraining its structure to just the necessary parts from the start.

Meng: That sounds like you’re trying to stop models from always trying to be exhaustive, which is what usually makes them slow and expensive at scale.

Tom: Exactly. The core idea is reducing that reasoning budget first so you can build on top of it later, making it schedulable >

Jane: And they call this Draft-Thinking Formulation because it cuts the initial reasoning budget down to set a baseline for efficiency >

Lu: They also take that efficient structure and distill some of those patterns from much bigger models, like DeepSeekV3-three hundred twenty-four to create a starting dataset for fine-tuning >

Tom: So they’re using distillation to get those high-quality draft reasoning examples without needing the full training cost of the massive model >

Jane: They then use progressive curriculum learning, which is kind of like a student learning: first mastering the basics through supervised fine-tuning and distillation, then gradually practicing harder tasks through reinforcement learning >

Meng: I’m watching that RL part. If it’s just another long chain of thought with a length penalty, it won't solve the efficiency problem >

Lu: No, they are using an incremental length expansion strategy to slowly increase the maximum generation length across those RL stages >

Tom: Because they found that if you start too long in the reinforcement learning phase, the model just reverts back to its old long chain of thought behavior, even with a draft prompt >

Jane: So by expanding the maximum output length slowly, they can build up that deep reasoning capability while keeping that concise draft form stable >

Meng: That’s smart. It means you can get the depth you need for a tough problem without immediately paying for it in tokens >

Lu: And at inference time, they give you a choice: either use the Draft mode to maximize token efficiency or switch to long chain of thought mode to maximize accuracy >

Tom: Or they have this instance adaptive prompt that looks at the question and decides whether to use one or the other dynamically based on difficulty >

Jane: It means you get a flexible system where the model handles simple questions quickly and switches to deep reasoning only when it's necessary for harder ones >

Meng: I’m still curious about the AIME two thousand twenty-five results. If this works well for everything, we need to know how it performs on those truly hard competition problems >

Lu: They actually found that while the efficiency is great across many benchmarks, they saw a bigger accuracy gap between the draft and long chain of thought modes on some of the most difficult datasets >

Tom: They suspect that those super complex problems might just require a longer maximum sequence length during the draft reasoning training phase to get the best results >

Jane: So while Draft-Thinking is very efficient, it seems like for the absolute hardest math competitions, you still need to give it more room to think >

Meng: It sounds like they’re trying to find that sweet spot where you have a model that’s smart enough but doesn't just overcommit its entire reasoning budget on every single question >

The paper's improvements: Tom: So, the paper outlines three main improvements they made to take Draft-Thinking from an idea to something that actually works on models >

Jane: Right, so they’re talking about how this method lets you have three different modes for using the AI at inference time >

Lu: That’s right. They have the standard draft mode which is all about token efficiency, and then you have the original long chain of thought mode which just maximizes accuracy >

Tom: But they also added this hybrid thing called instance adaptive prompting that lets the model check how hard a question is and switch between those two modes automatically >

Jane: It means for an easy question, it uses draft reasoning to save tokens, and if it sees a really tough one, it switches to the full chain of thought for better accuracy >

Meng: That dynamic switching is something I could actually deploy. It’s not just picking a fixed length; the AI decides the length based on what’s in front of it >

Lu: And they also added this concept of internalizing efficiency as an endogenous capability, so the model learns that concise draft structure itself >

Tom: That means it doesn't need an external routing system or some pre-set difficulty annotations to know which path to take >

Jane: It’s about building that efficient reasoning pattern right into the AI’s core ability during training >

Lu: And then they have this progressive curriculum learning strategy, which is like a careful training schedule that starts with supervised fine-tuning and distillation before moving into staged reinforcement learning >

Meng: That staged RL is still my main concern. It sounds like they use incremental length expansion to gradually increase the reasoning depth while trying to keep that efficient structure stable >

Tom: Yes, they found that if you start with a huge maximum output length at the beginning of the RL phase, the model just forgets what Draft-Thinking taught it about being concise >

Jane: So this staged approach lets them grow problem-solving capability while forcing it to maintain that concise reasoning form throughout >

Lu: They do this by using two stages with maximum output lengths of three thousand and then six thousand for their training on the AIME two thousand twenty-five dataset >

Tom: And the big win they’re showing is on the token efficiency numbers. On MATH five hundred, Draft-Thinking hits nine point one eight efficiency, which is much better than other RL methods like ThinkPrune >

Jane: So what this means for us listening right now is that if you’re running a model and you need to balance speed with correctness, this framework gives you a concrete way to do that >

Meng: I'm still watching the caveat on the AIME two thousand twenty-five set. They admitted the accuracy gap between draft and long chain of thought modes was bigger there than it was on other test sets >

Lu: They suspect that for those extremely hard problems, you might need to give the draft reasoning training a larger maximum sequence length to get closer to the full capability >

Tom: So if you’re an engineer building this, you gotta be aware that while it saves tokens generally, there’s still a potential ceiling on how deep it can go for the most complex tasks >

Conclusion: Tom: So we've gone through Draft-Thinking, learning how to teach models to internalize efficient reasoning structures >

Jane: Exactly, it’s about making efficiency an inherent part of the AI's thinking process so it doesn't waste tokens on unnecessary steps >

Lu: The main implication here is that we can schedule the reasoning budget based on how hard a task is, rather than having one fixed, long chain of thought for everything >

Meng: From a deployment standpoint, this means we can run models faster and cheaper because the inference cost scales better with the problem complexity >

Lalam: I’m quite certain that Draft-Thinking makes reasoning flexible and reliable for real-world applications because it learns the structure itself rather than relying on external prompts >

Tom: It’s about moving away from just forcing models to be long, to letting them choose their own depth based on the input >

Jane: And they show this in the results where Draft-Thinking achieved better token efficiency compared to methods like ThinkPrune across many benchmarks >

Lu: The Progressive Curriculum Learning strategy seems like a solid way to ensure that these efficient structures are learned stably, moving from supervised fine-tuning into staged reinforcement learning >

Meng: I’m still curious about the AIME two thousand twenty-five set. They flagged that the accuracy difference between draft and long chain of thought modes was bigger there than on other sets >

Tom: That suggests complex problems might still need a longer maximum sequence length during training to get peak accuracy, even with this method >

Jane: So while Draft-Thinking is great for efficiency across the board, it’s not a perfect solution for every single type of hard problem yet >

Lu: Overall, the work on Draft-Thinking shows how to internalize reasoning efficiency as an endogenous capability in large reasoning models >

Meng: It gives us a better tool for deploying AI that needs to be both smart and cheap at scale >

Lalam: I'm confident in this direction because it’s built on distillation and curriculum learning rather than just guessing the right prompt structure >

Jie Cao, Tianwei Lin, Zhenxuan Fan, Bo Yuan, Ziyuan Zhao, Rolan Yan, †Wenqiao Zhang, №Siliang Tang

Zhejiang University · Tencent

cs.AI, cs.CL

Submitted: 2026-02-28

Updated: 2026-10-14

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 89/100

The gist: The gist The Draft-Thinking method guides large reasoning models to first learn a concise draft-style reasoning structure that retains only the critical reasoning steps, and then stably internalizes

Key concepts

Draft-Thinking Formulation: This is the core idea of constraining the model's reasoning budget. Instead of always generating a full step-by-step trace, it learns to generate a compact 'draft' output that only includes essential reasoning steps. This forces the model to focus on high-value inferences immediately, setting a foundation for efficiency.
Progressive Curriculum Learning: This is the training strategy used to teach the model efficiently. It mimics student learning: first mastering core concepts via supervised fine-tuning (SFT) and distillation, then gradually internalizing these structures through two stages of reinforcement learning (RL). This ensures the model learns efficient reasoning before tackling complex tasks.
Incremental Length Expansion Strategy: This technique controls how deep the reasoning goes during RL. It starts with a relatively small maximum output length to prevent the model from reverting to its old, long CoT behavior. It then progressively increases this limit across training stages, allowing the model's reasoning depth to grow steadily while staying within an efficient budget.

Terminology

Core — 1× in this paper · explained in 2 episodes

Summary

The gist The Draft-Thinking method guides large reasoning models to first learn a concise draft-style reasoning structure that retains only the critical reasoning steps, and then stably internalizes this efficient pattern through progressive curriculum learning and adaptive prompting.

How it works

Draft-Thinking addresses the systematic issue of overthinking in long chain-of-thought (CoT) paradigms by guiding models to first learn a concise draft-style reasoning structure that retains only the critical reasoning steps, and then stably internalizes this efficient reasoning pattern as its capability scales Draft-Thinking, which guides models to first learn a concise draft-style reasoning structure that retains only the critical reasoning steps, and then stably internalizes this efficient reasoning pattern as its capability scales.

This is achieved through two main components:

Draft-Thinking Formulation: This formulation reduces the reasoning budget to improve efficiency and lays the foundation for subsequently elevating reasoning depth to a schedulable behavior Let M denote a reasoning large reasoning model (LRM) equipped with long chain-of-thought (CoT) capability. For any question q, under a “step-by-step” reasoning prompt pstep, the model generates a sampled output o ∼ M(pstep, q), where o = (rstep ⊕ a). The objective of Draft-Thinking is to learn, within the same model, a draft-style reasoning behavior constrained by key reasoning, such that under a draft prompt pdraft, the model generates o ∼ M(pdraft, q), where o = (rdraft ⊕ a) This formulation reduces the reasoning budget to improve efficiency and lays the foundation for subsequently elevating reasoning depth to a schedulable behavior.
Cultivating Draft Reasoning Capability: To develop draft reasoning capability for model M, we distill this ability from the larger model DeepSeekV3-0324 (685B), using it to generate high-quality draft reasoning data for SFT We start with LIMO (Ye et al., 2025), a curated mathematical dataset D = 1 with N = 817, selected from tens of millions of problems for its diverse difficulty, generality, and knowledge coverage. Using a carefully designed mathematical reasoning prompt method Chunked Symbolism (Aytes et al., 2025), we generate draft reasoning rdraft and answers aˆ for D Filtering samples where aˆ!= a yields our SFT dataset Dsft = 1. Subsequently, the distillation objective is to finetune model M with dataset Dsft to enable it to develop concise draft reasoning capability by minimizing the cross-entropy loss: L = − X l i=1 logPθM (yi pdraft, q, y<i) (1)

Progressive Curriculum Learning

To stably acquire the above efficient reasoning and scheduling capabilities, we draw intuitive motivation from the student learning process: to develop concise and reliable problem-solving strategies, students typically first master core concepts and inferences and then internalize them into stable problem-solving cognition through repeated practice Based on this insight, we propose a Progressive Curriculum Learning as illustrated in Figure 1 The model first fits draft-style reasoning structures at the behavioral level through supervised fine-tuning (SFT), and leverages black-box distillation to extract implicit structured inference patterns from stronger models, establishing a reliable low-budget reasoning framework It then introduces a two-stage reinforcement learning (RL) process to gradually expand problem-solving capability, stably learning hard tasks that require deep reasoning while maintaining a preference for lowbudget efficiency This training procedure does not rely on predefined compression ratios, templates, or length penalties; instead, it encourages the model to internalize the essential structure of concise reasoning through optimization

Enhancing Draft Reasoning Via Iterative Reinforcement Learning

We introduce an Incremental Length Expansion Strategy, which progressively relaxes the maximum generation length Lmax across multiple stages of reinforcement learning, allowing the model to steadily increase reasoning depth under a controlled budget This design is motivated by a key observation: imposing a large Lmax at the initial reinforcement learning stage causes the model to revert to its original long CoT behavior to obtain higher rewards, even under the draft prompt pdraft, thereby undermining the learned concise reasoning structure In contrast, stage-wise expansion of Lmax enables progressive improvement in problem-solving capability while preserving the draft reasoning form Specifically, we perform two-stage RL training with maximum output lengths Lmax of 3000 and 6000, respectively For the first stage, we use the 475 LIMO problems that remain after excluding the 342 samples in Dsft from the total 817 LIMO problems, forming Drl = 1. These problems are more challenging than those used in the SFT stage For the second stage, we use the more challenging AIME24 dataset, which contains 30 hard mathematical competition problems In both RL training stages, we use the Group Relative Policy Optimization (GRPO) (Shao et al., 2024) algorithm

Inference and Adaptive Prompting

At inference, the model can perform draft reasoning with prompt pdraft or its original long CoT reasoning with pstep Draft mode maximizes token efficiency while long CoT mode maximizes accuracy We also design a hybrid instance adaptive prompt padaptive that combines both modes, enabling the model to balance token efficiency with high accuracy by leveraging the strengths of each approach Specifically, for each question q, the model assesses difficulty and adaptively chooses between draft reasoning mode for simpler questions and long CoT mode for more challenging questions

Experimental Results

Draft-Thinking achieves significantly superior token efficiency (EFF) compared to all other methods As shown in Tables 1-2, Draft-Thinking attains the highest EFF across all mathematical reasoning benchmarks For instance, on MATH500, Draft-Thinking achieves an EFF of 9.18, substantially outperforming other RL methods on the same base model (e.g., ThinkPrune and DR.SAF), and even surpassing RL-based (O1-Pruner and ThinkPrune) and offline (SimPO(FCS+Reflection)) methods on larger 32B models Draft-Thinking simultaneously enhances the model’s long-CoT accuracy while substantially boosting its token efficiency After three stages of Draft-Thinking training, the model exhibits improved long CoT reasoning (step prompt) performance across all benchmarks while demonstrating substantial reductions in token length relative to the original baseline

Conclusion and Limitations

In summary, our key contributions are: (i) We propose Draft-Thinking, which internalizes the selection of high-value steps as an endogenous capability and reasoning pattern of LRMs, significantly outperforming existing methods in reasoning efficiency and flexibility (ii) We elevate reasoning depth to a decidable dimension, enabling the reasoning budget to be flexibly scheduled according to task complexity (iii) A Progressive Curriculum Learning strategy is introduced to integrate SFT with staged RL training, stably achieving the co-evolution of efficient reasoning structures and deep reasoning capability However, on the most difficult AIME 2025 dataset, the accuracy gap between draft and long-CoT modes is larger than on other sets We suspect that complex problems require a larger maximum sequence length during draft reasoning training

Evaluation Details

We follow LIMO’s comprehensive evaluation framework to evaluate the effectiveness of our method Our evaluation datasets include in-domain datasets (MATH500 (Hendrycks et al., 2021), AIME2025) and outof-distribution datasets (LiveMathBench (version 202505) (Liu et al., 2024), OlympiadBench (He et al., 2024), MinervaMath (Lewkowycz et al., 2011), GPQA (Rein et al.

Improvements for AI systems

The Draft-Thinking mechanism internalizes reasoning efficiency as an endogenous capability, allowing models to learn a concise draft-style reasoning structure that retains only the critical reasoning steps. This enables the model to switch between low-budget and more exhaustive reasoning without external routing or explicit difficulty annotations, thereby avoiding worst-case uniform expansion under heterogeneous task difficulty.
The Progressive Curriculum Learning strategy ensures stable acquisition of efficient reasoning by first fitting draft-style reasoning structures at the behavioral level through supervised fine-tuning (SFT), and leverages black-box distillation to extract implicit structured inference patterns from stronger models. This creates a reliable low-budget framework before introducing staged RL to gradually expand problem-solving capability, stably learning hard tasks that require deep reasoning while maintaining a preference for lowbudget efficiency.
The system introduces an adaptive prompting mechanism that allows the model to elevate reasoning depth to a flexible, model-selectable behavior by assessing task difficulty and choosing adaptively between draft reasoning mode for simpler questions and long CoT mode for more challenging questions, thereby achieving higher accuracy and higher EFF than long CoT reasoning on OlympiadBench and MATH500.
The inference phase allows the model to utilize three distinct modes: Draft mode maximizes token efficiency while long CoT mode maximizes accuracy, or a hybrid instance adaptive prompt that combines both modes, enabling the model to balance token efficiency with high accuracy by leveraging the strengths of each approach.

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 Draft-Thinking, a lightweight adaptation framework that learns a model-adapted draft CoT: a compact, self-contained working derivation that directly supports the final answer. We identify a Cross-Scale Draft Mismatch: concise traces generated by a strong teacher can be too sparse for a smaller target model. To address this mismatch, 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.