Elamur Research⌕
☰
home / papers / 2026 / Dissecting Failure Dynamics in Large Language Model Reasoning

Dissecting Failure Dynamics in Large Language Model Reasoning

Paper · 2026 · Wei Zhu, Jian Zhang, Lixing Yu et al.

Summary

Reasoning-oriented LLMs often improve by spending more compute at inference time, yet little is known about where in a single chain of thought things actually break. The authors study this using reasoning traces from DeepSeek-R1-Distill-Qwen-1.5B on AMC and AIME. They split each trace into segments and have each segment judged by a Gemini 3 Pro oracle, with human spot-checks. They find that mistakes usually begin at a few early turning points. After such a point, the model keeps producing plausible-looking steps built on the faulty premise, a pattern they call an epistemic spiral. These turning points line up with sudden jumps in token entropy, and resampling from the same prefix can still reach the right answer. Building on this, they propose GUARD, an inference-time method that follows one main trajectory. When entropy exceeds an adaptive, history-based percentile, GUARD briefly branches using Momentum, Inhibitory and Counterfactual continuations, and it also curbs drawn-out deliberation near the end. The approach improves accuracy without retraining and works across different backbones.

Key points

  • More than 85% of failure onsets fall within the first 30% of a trajectory, and 43.5% of faulty traces contain only one invalid segment.
  • Wrong trajectories are longer than correct ones, with a median of 14 versus 9 segments (Mann-Whitney U, p < 0.001).
  • On Llama-3.1-8B-Instruct, GUARD reaches a 26.7 average, compared with 19.6 for greedy decoding and 21.7 for EM-INF. It also helps the math-specialized JustRL-1.5B, including on coding tasks.
  • Ablation on DeepSeek-R1-Distill-Qwen-7B (average over eight benchmarks) gives 59.8 for the full method. Removing late-stage control drops this to 53.0, and removing Momentum, Counterfactual or Inhibitory branching drops it to 54.3, 55.4 and 57.1 respectively.
  • Code is released at github.com/ZHUWEI-hub/GUARD. Stated limitations are the reliance on an external oracle, the use of entropy as the only uncertainty signal, and an evaluation focused on structured benchmarks.

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

🎓Yunnan UniversityKunming · China20 works · 40 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksReasoning · 1,685 works

Cite

Cite as Wei Zhu, Jian Zhang, Lixing Yu, Kun Yue and Zhiwen Tang. 2026. Dissecting Failure Dynamics in Large Language Model Reasoning. 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.401

@inproceedings{zhu2026dissecting,
  title = {{Dissecting Failure Dynamics in Large Language Model Reasoning}},
  author = {Wei Zhu, Jian Zhang, Lixing Yu, Kun Yue and Zhiwen Tang},
  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.401},
  url = {https://doi.org/10.18653/v1/2026.acl-long.401}
}

References

55 references, 21 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Kejia Chen, Jiawen Zhang, Xiao Pan, Yihong Wu, Zunlei Feng, Mingli Song, and Ruoxi Jia. 2026a. CASPO: Confidence-aware step-wise preference optimization for reliable reasoning in large language models. find
  2. Xi Chen, Wei Xue, and Yike Guo. 2026b. Actormind: Emulating human actor reasoning for speech role-playing. Preprint, arXiv:2604.11103. find
  3. Xi Chen and Min Zeng. 2025. Prototype conditioned generative replay for continual learning in NLP. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL 2025 - Volume 1: Long Papers, Albuquerq… find
  4. Zhuokun Chen, Zeren Chen, Jiahao He, Lu Sheng, Mingkui Tan, Jianfei Cai, and Bohan Zhuang. 2025. R-stitch: Dynamic trajectory stitching for efficient reasoning. arXiv preprint arXiv:2507.17307. — [arXiv:2507.17307](https://arxiv.org/abs/2507.17307) find
  5. DeepSeek-AI. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. CoRR, abs/2501.12948.
  6. Tianyu Fu, Yi Ge, Yichen You, Enshu Liu, Zhihang Yuan, Guohao Dai, Shengen Yan, Huazhong Yang, and Yu Wang. 2025. R2r: Efficiently navigating divergent reasoning paths with small-large model token routing. arXiv preprint arXiv:2505.21600. — [arXiv:2505.21600](https://arxiv.org/abs/2505.21600) find
  7. Marah I Abdin, Jyoti Aneja, Harkirat S. Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J. Hewett, Mojan Javaheripi, Piero Kauffmann, James R. Lee, Yin Tat Lee, Yuanzhi Li, Weishung Liu, Caio C. T. Mendes, Anh Nguyen, Eric Price, Gustavo de Rosa, Olli Saarikivi, and… — Phi-4 Technical Report
  8. Google DeepMind. 2025. A new era of intelligence with gemini 3. https://blog.google/products/gemini/gemini-3/. Accessed: 2025-11-29. ↗
  9. Shivam Agarwal, Zimin Zhang, Lifan Yuan, Jiawei Han, and Hao Peng. 2025. The unreasonable effectiveness of entropy minimization in llm reasoning. arXiv preprint arXiv:2505.15134. find
  10. Bingxiang He, Zekai Qu, Zeyuan Liu, Yinghao Chen, Yuxin Zuo, Cheng Qian, Kaiyan Zhang, Weize Chen, Chaojun Xiao, Ganqu Cui, Ning Ding, and Zhiyuan Liu. 2025. Justrl: Scaling a 1.5b LLM with a simple RL recipe. CoRR, abs/2512.16649. — [arXiv:2512.16649](https://arxiv.org/abs/2512.16649) find
  11. AI-MO. 2024. AIMO Validation Dataset - AMC. https://huggingface.co/datasets/AI-MO/aimo-v alidation-amc. Accessed: 2024-11-19. ↗
  12. Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Zirun Guo, Minjie Hong, Feng Zhang, Kai Jia, and Tao Jin. 2025. Thinking with programming vision: Towards a unified view for thinking with images. Preprint, arXiv:2512.03746. — [arXiv:2512.03746](https://arxiv.org/abs/2512.03746) find
  13. Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Leng Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, Jie Liu, Lei Qi, Zhiyuan Liu, and Maosong Sun. 2024. Olympiadbench: A challenging benchmark for promoting AGI with 8901 9 olympiad-level bilingual multimodal scientific proble… find
  14. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874.
  15. Minjie Hong, Zetong Zhou, Zirun Guo, Ziang Zhang, Ruofan Hu, Weinan Gan, Jieming Zhu, and Zhou Zhao. 2025. Generative reasoning recommendation via llms. Preprint, arXiv:2510.20815. — [arXiv:2510.20815](https://arxiv.org/abs/2510.20815) find
  16. Jiameng Huang, Baijiong Lin, Guhao Feng, Jierun Chen, Di He, and Lu Hou. 2026. Efficient reasoning for large reasoning language models via certainty-guided reflection suppression. In Fortieth AAAI Conference on Artificial Intelligence, Thirty-Eighth Conference on Innovative Applications of Artificia… find
  17. Xiping Li and Jianghong Ma. 2025. Aimcot: Active information-driven multimodal chain-of-thought for vision-language reasoning. CoRR, abs/2509.25699. — [arXiv:2509.25699](https://arxiv.org/abs/2509.25699) find
  18. Shuquan Lian, Juncheng Liu, Yazhe Chen, Yuhong Chen, and Hui Li. 2026. Swe-agile: A software agent framework for efficiently managing dynamic reasoning context. Preprint, arXiv:2604.11716. find
  19. Baohao Liao, Yuhui Xu, Hanze Dong, Junnan Li, Christof Monz, Silvio Savarese, Doyen Sahoo, and Caiming Xiong. 2025. Reward-guided speculative decoding for efficient llm reasoning. arXiv preprint arXiv:2501.19324. — [arXiv:2501.19324](https://arxiv.org/abs/2501.19324) find
  20. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s verify step by step. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. Open-Rev…
  21. Guoming Ling, Zhongzhan Huang, Yupei Lin, Junxin Li, Shanshan Zhong, Hefeng Wu, and Liang Lin. 2026. Neural chain-of-thought search: Searching the optimal reasoning path to enhance large language models. Preprint, arXiv:2601.11340. find
  22. 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.
  23. Peiyang Liu, Xi Wang, Ziqiang Cui, and Wei Ye. 2025. Queries are not alone: Clustering text embeddings for video search. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2025, Padua, Italy, July 13-18, 2025, pages 874–883. ACM. find
  24. Yuxiang Ji, Yong Wang, Ziyu Ma, Yiming Hu, Hailang Huang, Xuecai Hu, Guanhua Chen, Liaoni Wu, and Xiangxiang Chu. 2026. Thinking with map: Reinforced parallel map-augmented agent for geolocalization. Preprint, arXiv:2601.05432. find
  25. Peiyang Liu, Jinyu Yang, Lin Wang, Sen Wang, Yunlai Hao, and Huihui Bai. 2023. Retrieval-based unsupervised noisy label detection on text data. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 4099–4104. find
  26. Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay V. Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, Behnam Neyshabur, Guy Gur-Ari, and Vedant Misra. 2022. Solving quantitative reasoning problems with language models. In Advances in…
  27. Bo Li, Mingda Wang, Gexiang Fang, Shikun Zhang, and Wei Ye. 2026a. Retrieval as generation: A unified framework with self-triggered information planning. Preprint, arXiv:2604.11407.
  28. Xianzhi Li, Ethan Callanan, Abdellah Ghassel, and Xiaodan Zhu. 2026b. Entropy-gated branching for efficient test-time reasoning. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2026 - Volume 1: Long Papers, Rabat, Morocco, March 24… find
  29. Shichao Ma, Zhiyuan Ma, Ming Yang, Xiaofan Li, Xing Wu, Jintao Du, Yu Cheng, Weiqiang Wang, Qiliang Liu, Zhengyang Zhou, and Yang Wang. 2026. TSPO: breaking the double homogenization dilemma in multi-turn search policy optimization. CoRR, abs/2601.22776.
  30. MAA Committees. 2025. AIME Problems and Solutions. https://artofproblemsolving.com/wi ki/index.php/AIME_Problems_and_Solutions. Accessed: 2025-11-19. ↗
  31. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. Self-refine: Iterative refinement with… find
  32. Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel J. Candès, and Tatsunori Hashimoto. 2025. s1: Simple test-time scaling. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, EMNL…
  33. Shuaiyi Nie, Siyu Ding, Wenyuan Zhang, Linhao Yu, Tianmeng Yang, Yao Chen, Tingwen Liu, Weichong Yin, Yu Sun, and Hua Wu. 2026. Attnpo: Attention-guided process supervision for efficient reasoning. arXiv preprint arXiv:2602.09953.
  34. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processi…
  35. Fei Wu, Zhenrong Zhang, Qikai Chang, Jianshu Zhang, Quan Liu, and Jun Du. 2026. Step potential advantage estimation: Harnessing intermediate confidence and correctness for efficient mathematical reasoning. Preprint, arXiv:2601.03823.
  36. CoRR, Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025a. Chain of draft: Thinking faster by writing less. arXiv preprint arXiv:2502.18600.
  37. Qwen Team. 2025. QwQ-32B-Preview: Preview of Qwen QwQ-32B. https://qwenlm.github.io/b log/qwq-32b-preview/. Accessed: 2025-03-20. ↗
  38. Zicheng Xu, Guanchu Wang, Yu-Neng Chuang, Guangyao Zheng, Alexander S Szalay, Zirui Liu, and Vladimir Braverman. 2025b. Dts: Enhancing large reasoning models via decoding tree sketching. arXiv preprint arXiv:2511.00640. — [arXiv:2511.00640](https://arxiv.org/abs/2511.00640) find
  39. OpenAI. 2024. Openai o1 system card. abs/2412.16720.
  40. 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.
  41. Daniel Scalena, Leonidas Zotos, Elisabetta Fersini, Malvina Nissim, and Ahmet Üstün. 2025. Eager: Entropy-aware generation for adaptive inference-time scaling. arXiv preprint arXiv:2510.11170. — [arXiv:2510.11170](https://arxiv.org/abs/2510.11170) find
  42. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634–8652.
  43. Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In The Thirteenth International Conference on Learning Representations. find
  44. Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. 2022. Solving math word problems with process-and outcome-based feedback. arXiv preprint arXiv:2211.14275.
  45. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V. Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5,…
  46. Yi Wang, Junxiao Liu, Shimao Zhang, Jiajun Chen, and Shujian Huang. 2025. Pats: Process-level adaptive thinking mode switching. arXiv preprint arXiv:2505.19250. — [arXiv:2505.19250](https://arxiv.org/abs/2505.19250) find
  47. An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, and 43 others. 2024. Qwen2 technical report. CoRR, abs/2407.1067… find
  48. Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Zheng Lin, Li Cao, and Weiping Wang. 2025a. Dynamic early exit in reasoning models. CoRR, abs/2504.15895.
  49. Wang Yang, Xiang Yue, Vipin Chaudhary, and Xiaotian Han. 2025b. Speculative thinking: Enhancing small-model reasoning with large model guidance at inference time. arXiv preprint arXiv:2504.12329. — [arXiv:2504.12329](https://arxiv.org/abs/2504.12329) find
  50. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809–11822. find
  51. Xixian Yong, Xiao Zhou, Yingying Zhang, Jinlin Li, Yefeng Zheng, and Xian Wu. 2025. Think or not? exploring thinking efficiency in large reasoning models via an information-theoretic lens. arXiv preprint arXiv:2505.18237. — [arXiv:2505.18237](https://arxiv.org/abs/2505.18237) find
  52. Haohan Yuan and Haopeng Zhang. 2025. Understanding LLM reasoning for abstractive summarization. CoRR, abs/2512.03503.
  53. Jinghan Zhang, Xiting Wang, Fengran Mo, Yeyang Zhou, Wanfu Gao, and Kunpeng Liu. 2025a. Entropy-based exploration conduction for multi-step reasoning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 3895–3906, Vienna, Austria. Association for Computational Linguistics.… find
  54. Junyu Zhang, Runpei Dong, Han Wang, Xuying Ning, Haoran Geng, Peihao Li, Xialin He, Yutong Bai, Jitendra Malik, Saurabh Gupta, and Huan Zhang. 2025b. Alphaone: Reasoning models thinking slow and fast at test time. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Process…
  55. Wei Zhu, Zhiwen Tang, and Kun Yue. 2025. Symphony: Synergistic multi-agent planning with heterogeneous language model assembly. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. 8904 12 find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhu-2026-dissecting-failure-dynamics-large-language/