Beyond Token Length: Step Pruner for Efficient and Accurate Reasoning in Large Language Models
Paper · 2026 · Canhui Wu, Qiong Cao, Chang Li et al.
Summary
Large Reasoning Models often "overthink," producing very long chains of thought even for trivial questions. Earlier RL fixes penalize token counts, but fewer tokens do not always mean fewer reasoning steps, and late in training models can game such penalties by dropping the reasoning altogether. The authors propose Step Pruner (SP), an RL framework trained with GRPO.
SP works in four stages:
- Each sampled response is split into steps, with paragraphs as the unit.
- The smallest step count among correct samples serves as the target.
- Correct responses that use extra steps are penalized, while wrong responses receive no step reward.
- Training halts once average output length stops decreasing, which prevents hacking through step merging.
On AIME24, GPQA Diamond, MATH500 and GSM8K, using distilled DeepSeek-R1 Qwen (7B, 1.5B) and Llama-8B models, SP matches or improves accuracy while sharply cutting output length. Across these settings it gives the best accuracy–length trade-off against seven baselines, including O1-Pruner, ShorterBetter and TrainEfficient. The work suggests that counting steps, rather than tokens, is a more reliable signal for making reasoning cheaper without sacrificing quality.
Key points
- SP cuts the number of tokens it uses on AIME24 by 69.7%. On MATH500 it needs about 33% of the tokens of DeepSeek-R1-Distilled-Qwen-7B at the same accuracy level (92.0% vs. 91.8%).
- Averaged over benchmarks, SP uses roughly 44% of baseline tokens. The comparison covers seven baselines, among them CCoT, CoD, Cot-Valve, Model Merge, O1-Pruner, ShorterBetter and TrainEfficient.
- Among the segmentation strategies tested, paragraph-based splitting gave the best balance between compression and accuracy.
- An ablation removed four components: correct reward, correct optimal step, masking of wrong responses, and skipping all-wrong groups. Each removal hurt accuracy, length, or both.
- Gemini 2.5 was used to label sentences in model outputs. SP-7B outputs contain more Pivotal Reasoning and calculation, and less exploration of alternatives and self-verification. On Llama-3.1-8B, SP reaches 84.8% on MATH500 with shorter outputs.
Institutions
ChinaFields, topics and concepts
📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 works
🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksReasoning · 1,426 works
Cite
Cite as Canhui Wu, Qiong Cao, Chang Li, Zhenfang Wang, Chao Xue, Yuwei Fan, Wei Xi and Xiaodong He. 2026. Beyond Token Length: Step Pruner for Efficient and Accurate Reasoning in Large Language Models. Findings of the Association for Computational Linguistics: ACL 2026.
@inproceedings{wu2026beyond,
title = {{Beyond Token Length: Step Pruner for Efficient and Accurate Reasoning in Large Language Models}},
author = {Canhui Wu, Qiong Cao, Chang Li, Zhenfang Wang, Chao Xue, Yuwei Fan, Wei Xi and Xiaodong He},
booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
year = {2026},
doi = {10.18653/v1/2026.findings-acl.94},
url = {https://doi.org/10.18653/v1/2026.findings-acl.94}
}References
- Daman Arora and Andrea Zanette. 2025. Training language models to reason efficiently. Preprint, arXiv:2502.04463. \#\# Conclusion find
Updated
This page in Markdown → https://md.research.elamur.ai/free/en/papers/wu-2026-beyond-token-length-step-pruner/