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.
Figures & tables
Figure 1: Overview of Draft-Thinking. (A) Cross-Scale Draft Mismatch occurs when a strong teacher’s concise trace omits intermediate transitions required by a smaller target model. (B) Progressive Draft Adaptation initializes a concise reasoning form through SFT and then calibrates the draft’s reasoning granularity through two correctness-driven RL stages: tight-budget adaptation and hard-task adaptation. (C) The adapted model supports draft, adaptive, and step-by-step prompts as parallel inference modes that prioritize token efficiency, provide a balanced trade-off, and elicit detailed long CoT reasoning, respectively.
Method
MATH500
AIME2025
LiveMathBench
OlympiadBench
Acc
Tokens
EFF
Acc
Tokens
EFF
Acc
Tokens
EFF
Acc
Tokens
EFF
Qwen3-8B
Original
93.00
5668
1.64
64.38
18034
0.36
32.17
17493
0.18
64.44
11887
0.54
SFT-only
91.00
4364
2.08
49.58
13129
0.37
26.17
13160
0.19
57.33
10424
0.55
GRPO-only
93.60
3074
3.04
53.96
11720
0.46
28.67
11043
0.26
65.48
6902
0.94
ThinkPrune
94.40
2615
3.61
48.13
9781
0.49
28.67
9333
0.31
65.63
5809
1.13
Table 1: Controlled comparison of draft-mode reasoning efficiency on four mathematical reasoning benchmarks using Qwen3-8B. The highest and second-highest EFF values are shown in bold and underlined , respectively. Draft-Thinking is evaluated under the draft prompt pdraft .
Method
Minerva
GPQA-D
Acc
Tokens
EFF
Acc
Tokens
EFF
Qwen3-8B
Original
52.21
7640
0.68
55.05
11554
0.47
SFT-only
47.06
7349
0.64
37.37
16367
0.23
GRPO-only
53.68
3822
1.40
60.61
6322
0.96
ThinkPrune
54.04
2725
1.98
60.10
5166
1.16
Table 2: Controlled comparison on two transfer benchmarks using Qwen3-8B.
Method
Prompt
MATH500
AIME2025
LiveMathBench
OlympiadBench
Minerva
GPQA-D
Acc/Tokens
EFF
Acc/Tokens
EFF
Acc/Tokens
EFF
Acc/Tokens
EFF
Acc/Tokens
EFF
Acc/Tokens
EFF
Qwen3-8B
Original
step
93.00/5668
1.64
64.38 /18034
0.36
32.17 /17493
0.18
64.44/11887
0.54
52.21 /7640
0.68
55.05 /11554
0.47
Draft-Thinking
step
93.60 /4271
2.19
64.79 /14112
0.46
30.83/13980
0.22
66.37 /9519
0.69
53.68 /5180
1.03
63.13 /6620
0.95
adaptive
93.20 /2755
3.38
56.67/12155
0.46
31.83 /10341
0.31
69.19 /6849
1.01
48.53/2979
1.62
52.02/4800
1.08
draft
90.60/986
9.18
45.42/6527
0.69
22.33/6476
0.35
61.19/3433
1.78
46.32/960
4.82
54.04/2689
2.01
Table 3: Comparison of three prompt-controlled inference modes of the same adapted Qwen3-8B model. The draft, adaptive, and step-by-step rows share identical model parameters and differ only in the inference prompt.
Method
Prompt
MATH500
AIME2025
LiveMathBench
OlympiadBench
Minerva
GPQA-D
Acc/Tokens
EFF
Acc/Tokens
EFF
Acc/Tokens
EFF
Acc/Tokens
EFF
Acc/Tokens
EFF
Acc/Tokens
EFF
Qwen3-8B
Original
step
93.00/5668
1.64
64.38/18034
0.36
32.17/17493
0.18
64.44/11887
0.54
52.21/7640
0.68
55.05/11554
0.47
Original
draft
95.00/3714
2.55
65.62/16254
0.40
27.00/17990
0.15
65.33/9968
0.65
51.47/4889
1.05
57.07/8143
0.70
D-SFT
draft
69.80/1732
4.02
9.17/3928
0.23
12.50/3732
0.33
36.89/3535
1.04
37.87/549
6.88
41.92/1948
2.15
D-SFT->D-RL6k
draft
93.00/1453
6.39
48.75/7426
0.65
26.33/6949
0.37
63.41/4340
1.46
49.26/1677
2.93
57.07/3742
1.52
Table 4: Ablation study of Draft-Thinking on Qwen3-8B. Original (step) is the unadapted long CoT reference, whereas Original (draft) applies pdraft without training. D-SFT uses draft-form SFT only; D-SFT → D-RL6k omits tight-budget adaptation and performs all 120 RL steps with Lmax=6,000 ; D-SFT → D-RL3k (60 steps) is the intermediate checkpoint after tight-budget adaptation; and Draft-Thinking is the full D-SFT → D-RL3k (60 steps) → D-RL6k (60 steps) pipeline.
Figure 2: Reasoning behavior comparison between original Qwen3-8B and Draft-Thinking on MATH500. Each bar represents the cumulative number of reasoning steps within a phase category.
Metric
Original
Draft
Step
Adaptive
Avg Steps
111
22
88
50
Avg Tokens per Step
51
45
48
55
Table 5: Average number of reasoning steps and average tokens per step on MATH500. We use double newlines (" \n\n ") as a step delimiter.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Prompt Method
Content
Long CoT
Please reason step by step, and put your final answer within \boxed{}.
Draft
Let’s think step by step, but only keep a minimum draft for each thinking step, with as few words as possible, and output the final answer within \boxed{}.
Instance adaptive
First, you should decide the thinking mode based on the question’s difficulty. The first mode is the normal way, where you can think in detail. The second mode is to keep only a minimal draft with as few words as possible. Then reason step by step, and output the final answer within \boxed{}.
Appendix
Table 6: Detailed designs of three prompts: pstep for long CoT reasoning, pdraft for draft reasoning, and padaptive for instance-adaptive reasoning.
Figure 3: Example resoning trace of original Qwen3-8B and Draft thinking on a MATH500 question.
Parameter
D-RL3k
D-RL6k
datasets
Drl
AIME2024
advantage estimator
grpo
train batch size
96
30
max prompt length
240
500
max response length
3000
6000
actor optim lr
1e-6
Appendix
Table 7: Main RL training configurations of Draft-Thinking on Qwen3-8B using the Verl framework. D-RL3k and D-RL6k denote the two progressive RL stages with maximum response lengths of 3,000 and 6,000 tokens, respectively.
Parameter
D-RL3k
D-RL6k
datasets
Drl
AIME2024
advantage estimator
grpo
train batch size
96
30
max prompt length
240
500
max response length
3000
6000
actor optim lr
1e-6
Appendix
Table 8: Main RL training configurations of Draft-Thinking on Qwen3-4B using the Verl framework. D-RL3k and D-RL6k denote the two progressive RL stages with maximum response lengths of 3,000 and 6,000 tokens, respectively.
Parameter
SFT-only
Draft-SFT (8B)
Draft-SFT (4B)
Base model
Qwen3-8B
Qwen3-8B
Qwen3-4B
Training data
LIMO
Drl
Drl
Training samples
817
342
342
Max sequence length
10,000
6,000
6,000
Global batch size
60
60
60
Micro batch size/GPU
1
1
2
Appendix
Table 9: SFT training configurations. Draft-SFT uses the filtered DeepSeek-V3-generated draft reasoning dataset Drl , while SFT-only uses the original LIMO solutions. The maximum sequence length refers to the combined prompt and response sequence.
Method
Prompt
Minerva
AIME2025
LiveMathBench
MATH500
Acc
Tokens
EFF
Acc
Tokens
EFF
Acc
Tokens
EFF
Acc
Tokens
EFF
Qwen3-4B
Original
step
51.47
7253
0.71
63.12
17703
0.36
29
16813
0.17
92.2
5505
1.67
D-SFT->D-RL3k->D-RL6k
draft
46.69
1421
3.28
46.04
7097
0.65
22.67
6051
0.38
92
1269
7.24
step
48.9
4815
1.01
60.21
12885
0.46
26.67
12863
0.21
92.8
4130
2.24
adaptive
48.9
3620
1.35
53.33
10048
0.53
24.17
8353
0.29
92.8
2292
4.04
Appendix
Table 10: Comparison of different training strategies on Qwen3-4B.
Figure 4: Comparison of average accuracy and response length across different difficulty levels on MATH500.
Figure 5: Average lengths of correct vs. wrong responses on MATH500.
Figure 6: Accuracy, token count, and token efficiency trends for three reasoning modes (draft, long CoT, and adaptive) across the three training stages of Draft-Thinking on Qwen3-8B.
Figure 7: Accuracy, token count, and token efficiency trends for three reasoning modes (draft, long CoT, and adaptive) across the three training stages of Draft Thinking on Qwen3-4B.
Figure 8: Chunked Symbolism Prompt Aytes et al. (2025) .
Large Language Models (LLMs) have achieved remarkable success in complex reasoning tasks through Chain-of-Thought (CoT) prompting. However, these models often exhibit "computational overthinking," generating redundant reasoning steps that increase latency and cost without improving accuracy. Recent studies suggest that CoT trajectories can be significantly pruned, yet existing methods often rely on forcing a static thinking budget, heuristic filtering, sub-optimal early exit via classification, or expensive re-training. In this paper, we introduce OS-Pruner, a lightweight plug-in framework that formulates chain-of-thought pruning as an optimal stopping problem. Given a reasoning prefix, OS-Pruner learns whether further reasoning is worth its token cost by optimizing an explicit utility that trades off final-answer accuracy against generated length. Our novel formulation enables the model to dynamically assess the sufficient point of termination for a reasoning chain. OS-Pruner is designed to be lightweight during both training and inference, and to provide users with fine-grained control over the reasoning-effort vs. accuracy trade-off. On diverse reasoning benchmarks and base models, OS-Pruner achieves 20-60% reduction in generation length with minimal accuracy sacrifice.
Mohammed Ehab, Aymane El Gadarri, Vivek F. Farias +2
Explicit chain-of-thought (CoT) reasoning substantially improves the reasoning ability of large language models (LLMs), but incurs high inference cost due to lengthy autoregressive traces. Existing latent reasoning methods offer a promis- ing alternative, yet they often treat reasoning as uniformly compressible, causing precision-critical intermediate steps to be overly compressed and thereby degrading reasoning accuracy. In this work, we propose Selective Latent Thinking (SLT), a reasoning paradigm that selectively compresses redundant reasoning spans into latent representations while preserving precision-critical spans as explicit CoT within the same reasoning trajectory. Our primary instantiation uses a lightweight decoder to anticipate a short upcoming reasoning span and applies confidence- based gating to determine the longest span that can be reliably compressed; the resulting selective compression policy is further optimized with trajectory-level reinforcement learning to balance answer correctness against reasoning cost. Ex- periments across four mathematical reasoning benchmarks show that which spans are compressed is critical: with the same compressor, compressing arbitrary spans reduces the GSM-Hard accuracy of Qwen3-4B from 54.7% to 17.3%, whereas keeping numerical tokens explicit retains 51.8%. SLT reduces explicit reasoning length by 58.4% with only a 2.8-point accuracy drop compared to explicit CoT on Llama-3.2-1B. Notably, the principle extends beyond this design: instantiated within existing latent reasoning models via a learned compression gate, it improves accuracy at equal reasoning length (+2.6 points over Latent-SFT) and at shorter length (+5.6 points over Latent-GRPO with 13% fewer tokens).
Hui Xie, Jie Liu, Ziyue Qiao +1
Eindhoven University of Technology, Netherlands · School of Computing and Information Technology, Great Bay University, China
Large reasoning models (LRMs) like OpenAI o1 and DeepSeek-R1 achieve high accuracy on complex tasks by adopting long chain-of-thought (CoT) reasoning paths. However, the inherent verbosity of these processes frequently results in redundancy and overthinking. To address this issue, existing works leverage Group Relative Policy Optimization (GRPO) to reduce LRM output length, but their static length reward design cannot dynamically adapt according to the relative problem difficulty and response length distribution, causing over-compression and compromised accuracy. Therefore, we propose SmartThinker, a novel GRPO-based efficient reasoning method with progressive CoT length calibration. SmartThinker makes a two-fold contribution: First, it dynamically estimates the optimal length with peak accuracy during training and guides overlong responses toward it to reduce response length while sustaining accuracy. Second, it dynamically modulates the length reward coefficient to avoid the unwarranted penalization of correct reasoning paths. Extensive experiment results show that SmartThinker achieves up to 52.5% average length compression with improved accuracy, and achieves up to 16.6% accuracy improvement on challenging benchmarks like AIME25. The source code can be found at https://github.com/SJTU-RTEAS/SmartThinker.
Chenzhi Hu, Qinzhe Hu, Yuhang Xu +6
Shanghai Jiao Tong University · Beihang University