Elamur Research⌕
☰
home / papers / 2026 / Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention

Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention

Paper · 2026 · Zhen Yang, Mingyang Zhang, Feng Chen et al.

Summary

Test-time scaling techniques such as self-consistency or tree search make LLMs reason better, but they pay for it with heavy extra inference compute. The authors analysed AIME2024 trajectories and found that wrong answers carry more total entropy than correct ones. The gap comes mostly from a small number of highly uncertain tokens, where errors seem to start and then spread. Based on this, they propose Minimal Test-Time Intervention (MTI), a training-free decoding scheme with two parts:

  • Selective CFG: classifier-free guidance is triggered only at positions where token entropy exceeds a threshold τ.
  • Cheap negative prompt: instead of keeping a full separate unconditional branch, the phrase "OUTPUT ERROR" is injected while reusing the main model's KV cache. This costs about two extra tokens per intervention.

MTI improves results on general, coding and STEM benchmarks across several recent LLMs and also transfers to a vision-language model. Its latency often stays close to plain decoding, partly because responses get shorter. The work matters because it shows that reasoning reliability can improve without heavy compute, through a plug-and-play method that works with existing acceleration tools.

Key points

  • Main gains: +9.28% average over six benchmarks for DeepSeek-R1-7B, and +11.25% on AIME2024 with Ling-mini-2.0.
  • Cost comparison (Ling-mini-2.0, total over GPQA-Diamond, Math500, HumanEval, Winogrande and HumanEvalPlus):
    • Direct inference: 78.26% accuracy, 4547 s.
    • Vanilla CFG: 76.30% accuracy, 12,411 s.
    • MTI with τ=1.0: 80.70% accuracy, 4551 s.
  • Why selectivity works: a greedy-search analysis with Qwen3-8B on HumanEvalPlus showed that CFG seldom changes low-entropy tokens. It mainly flips predictions where the model is uncertain.
  • Vision-language transfer: with Qwen3-8B-VL-Instruct at τ=2.0, guidance fired on only 3.7–4.5% of tokens. Accuracy rose on Where2Place (+2.97), RefSpatial (+1.79) and VaBenchPoint (+1.99).
  • Open question: meaningless prompts (e.g., "apple") also beat direct inference, though less than "OUTPUT ERROR". The authors note this effect is not yet explained.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Hong Kong University of Science and TechnologyHong Kong · China142 works · 25,136 citations🎓Zhejiang UniversityHangzhou · China172 works · 12,988 citations🏢Ant GroupHangzhou · China42 works · 1,131 citations🎓The Hong Kong University of Science and Technology (Guangzhou)Guangzhou · China53 works · 1,367 citations🏢KuaishouBeijing · China15 works · 95 citations

 Fields, 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 Zhen Yang, Mingyang Zhang, Feng Chen, Ganggui Ding, Liang Hou, Xin Tao and Ying-Cong Chen. 2026. Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.921

@inproceedings{yang2026less,
  title = {{Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention}},
  author = {Zhen Yang, Mingyang Zhang, Feng Chen, Ganggui Ding, Liang Hou, Xin Tao and Ying-Cong Chen},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.921},
  url = {https://doi.org/10.18653/v1/2026.acl-long.921}
}

References

52 references, 23 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Pondé de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, and 39 others. 202… — Evaluating Large Language Models Trained on Code
  2. Mouxiang Chen, Binyuan Hui, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Jianling Sun, Junyang Lin, and Zhongxin Liu. 2025a. Parallel scaling law for language models. arXiv preprint arXiv:2505.10475. — [arXiv:2505.10475](https://arxiv.org/abs/2505.10475) find
  3. Zhuokun Chen, Zeren Chen, Jiahao He, Lu Sheng, Mingkui Tan, Jianfei Cai, and Bohan Zhuang. 2025b. R-stitch: Dynamic trajectory stitching for efficient reasoning. arXiv preprint arXiv:2507.17307. — [arXiv:2507.17307](https://arxiv.org/abs/2507.17307) find
  4. Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James Glass, and Pengcheng He. 2023. Dola: Decoding by contrasting layers improves factuality in large language models. arXiv preprint arXiv:2309.03883.
  5. OpenCompass Contributors. 2023. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/ opencompass. ↗
  6. Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems. find
  7. DeepSeek-AI. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. CoRR, abs/2501.12948.
  8. Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. arXiv preprint arXiv:1805.04833.
  9. Yichao Fu, Junda Chen, Siqi Zhu, Zheyu Fu, Zhongdongming Dai, Aurick Qiao, and Hao Zhang. 2024. Efficiently serving LLM reasoning programs with certaindex. CoRR, abs/2412.20993. — [arXiv:2412.20993](https://arxiv.org/abs/2412.20993) find
  10. Yichao Fu, Xuewei Wang, Yuandong Tian, and Jiawei Zhao. 2025. Deep think with confidence. arXiv preprint arXiv:2508.15260. find
  11. Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2025. Token-budget-aware llm reasoning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 24842–24855.
  12. Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598.
  13. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751.
  14. Jiameng Huang, Baijiong Lin, Guhao Feng, Jierun Chen, Di He, and Lu Hou. 2025. Efficient reasoning for large reasoning language models via certainty-guided reflection suppression. arXiv preprint arXiv:2508.05337. — [arXiv:2508.05337](https://arxiv.org/abs/2508.05337) find
  15. Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. 2024. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974.
  16. Felix Koulischer, Johannes Deleu, Gabriel Raya, Thomas Demeester, and Luca Ambrogioni. 2024. Dynamic negative guidance of diffusion models: Towards immediate content removal. In Neurips Safe Generative AI Workshop 2024. find
  17. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th symposium on operating systems principles, pages 611–626.
  18. Ayeong Lee, Ethan Che, and Tianyi Peng. 2025a. How well do llms compress their own chain-of-thought? a token complexity approach. arXiv preprint arXiv:2503.01141. — [arXiv:2503.01141](https://arxiv.org/abs/2503.01141) find
  19. Hakyung Lee, Subeen Park, Joowang Kim, Sungjun Lim, and Kyungwoo Song. 2025b. Uncertainty-aware contrastive decoding. In Findings of the Association for Computational Linguistics: ACL 2025, pages 26376–26391. find
  20. Yaniv Leviathan, Matan Kalman, and Yossi Matias. 2023. Fast inference from transformers via speculative decoding. In International Conference on Machine Learning, pages 19274–19286. PMLR. find
  21. Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori B Hashimoto, Luke Zettlemoyer, and Mike Lewis. 2023. Contrastive decoding: 20132 9 find
  22. Open-ended text generation as optimization. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers), pages 12286–12312. find
  23. Xianzhi Li, Ethan Callanan, Abdellah Ghassel, and Xiaodan Zhu. 2025a. Entropy-gated branching for efficient test-time reasoning. arXiv preprint arXiv:2503.21961. find
  24. Zeju Li, Jianyuan Zhong, Ziyang Zheng, Xiangyu Wen, Zhijian Xu, Yingying Cheng, Fan Zhang, and Qiang Xu. 2025b. Compressing chain-of-thought in llms via step entropy. arXiv preprint arXiv:2508.03346. — [arXiv:2508.03346](https://arxiv.org/abs/2508.03346) find
  25. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s verify step by step. In The Twelfth International Conference on Learning Representations.
  26. Anton Lozhkov, Raymond Li, Loubna Ben Allal, Federico Cassano, Joel Lamy-Poirier, Nouamane Tazi, Ao Tang, Dmytro Pykhtar, Jiawei Liu, Yuxiang Wei, Tianyang Liu, Max Tian, Denis Kocetkov, Arthur Zucker, Younes Belkada, Zijian Wang, Qian Liu, Dmitry Abulkhanov, Indraneil Paul, and 38 others. 2024. Sta… find
  27. Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. 2025. Reasoning models can be effective without thinking. arXiv preprint arXiv:2504.09858. find
  28. Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. 2025. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2063–2072. IEEE. find
  29. Minh Nhat Nguyen, Andrew Baker, Clement Neo, Allen Roush, Andreas Kirsch, and Ravid Shwartz-Ziv. 2024. Turning up the heat: Min-p sampling for creative and coherent llm outputs. arXiv preprint arXiv:2407.01082. — [arXiv:2407.01082](https://arxiv.org/abs/2407.01082) find
  30. OpenAI. 2023. GPT-4 technical report. abs/2303.08774.
  31. CoRR, Chen Qian, Dongrui Liu, Haochen Wen, Zhen Bai, Yong Liu, and Jing Shao. 2025. Demystifying reasoning dynamics with mutual information: Thinking tokens are information peaks in llm reasoning. arXiv preprint arXiv:2506.02867. — [arXiv:2506.02867](https://arxiv.org/abs/2506.02867) find
  32. Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy P. Lillicrap, Jean-Baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, Ioannis Antonoglou, Rohan Anil, Sebastian Borgeaud, Andrew M. Dai, Katie Millican, Ethan Dyer, Mia Glaese, Thibault Sotti… — Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context
  33. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. 2024. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling.
  34. Guillaume Sanchez, Honglu Fan, Alexander Spangher, Elad Levi, Pawan Sasanka Ammanamanchi, and Stella Biderman. 2023. Stay on topic with classifier-free guidance. arXiv preprint arXiv:2306.17806. — [arXiv:2306.17806](https://arxiv.org/abs/2306.17806) find
  35. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans,…
  36. Chenxia Tang, Jianchun Liu, Hongli Xu, and Liusheng Huang. 2025. Eliminating noise in logit space for robust token sampling of llm. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10758–10774. find
  37. Ling Team and Inclusion AI. 2025. Every activation boosted: Scaling general reasoner to 1 trillion open language foundation. CoRR, abs/2510.22115. — [arXiv:2510.22115](https://arxiv.org/abs/2510.22115) find
  38. Llama Team. 2024. The llama 3 herd of models. CoRR, abs/2407.21783.
  39. Qwen Team. 2025. Qwen3 technical report. CoRR, abs/2505.09388.
  40. Shenzhi Wang, Le Yu, Chang Gao, Chujie Zheng, Shixuan Liu, Rui Lu, Kai Dang, Xionghui Chen, Jianxin Yang, Zhenru Zhang, Yuqiong Liu, An Yang, Andrew Zhao, Yang Yue, Shiji Song, Bowen Yu, Gao Huang, and Junyang Lin. 2025. Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforceme…
  41. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.
  42. Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, Tianle Li, Max Ku, Kai Wang, Alex Zhuang, Rongqi Fan, Xiang Yue, and Wenhu Chen. 2024. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark…
  43. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. Preprint, arXiv:2201.11903. 20133 10
  44. Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, and Wenjie Li. 2025. Tokenskip: Controllable chain-of-thought compression in llms. arXiv preprint arXiv:2502.12067.
  45. gpqa_gen.py Minghao Yan, Saurabh Agarwal, and Shivaram Venkataraman. 2024. Decoding speculative decoding. arXiv preprint arXiv:2402.01528. find
  46. aime2024_gen.py • math_500_gen.py • humaneval_gen.py Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Qiaowei Li, Minghui Chen, Zheng Lin, and Weiping Wang. 2025. Dynamic early exit in reasoning models. arXiv preprint arXiv:2504.15895. find
  47. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Karthik Narasimhan, Mohit Hodosh, Eric Edwards, Jieyu Xu, Yuan Zhao, and Karthik Duan. 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems. find
  48. Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, Kaidong Yu, Peng Liu, Qiang Liu, Shawn Yue, Senbin Yang, Shiming Yang, Tao Yu, Wen Xie, Wenhao Huang, and 11 others. 2024. Yi: Open foundation models by 01.ai. CoRR, abs/2403.046…
  49. livecodebench_gen.py • mmlu_pro_gen.py During evaluation, the MMLU-Pro score is computed as the average over all subsets. For AIME2024, we conduct eight runs and report the mean score. The results of LiveCodeBench come from tests on lcb_code_generation and lcb_code_execution. A.2 Model Qwen3-8B Zhen… — [arXiv:2310.12149](https://arxiv.org/abs/2310.12149) find
  50. humaneval_plus_gen.py Wentao Yuan, Jiafei Duan, Valts Blukis, Wilbert Pumacay, Ranjay Krishna, Adithyavairavan Murali, Arsalan Mousavian, and Dieter Fox. 2024. Robopoint: A vision-language model for spatial affordance prediction for robotics. arXiv preprint arXiv:2406.10721. find
  51. Jun Zhang, Jue Wang, Huan Li, Lidan Shou, Ke Chen, Gang Chen, and Sharad Mehrotra. 2023. Draft & verify: Lossless large language model acceleration via self-speculative decoding. arXiv preprint arXiv:2309.08168.
  52. Enshen Zhou, Jingkun An, Cheng Chi, Yi Han, Shanyu Rong, Chi Zhang, Pengwei Wang, Zhongyuan Wang, Tie-Jun Huang, Lu Sheng, and Shanghang Zhang. 2025. Roborefer: Towards spatial referring with reasoning in vision-language models for robotics. CoRR, abs/2506.04308. — [arXiv:2506.04308](https://arxiv.org/abs/2506.04308) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yang-2026-less-is-more-improving-llm/