Elamur Research⌕
☰
home / papers / 2026 / The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis

The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis

Paper · 2026 · Zihao Wei, Liang Pang, Jiahao Liu et al.

Summary

Reasoning LLMs often keep generating "thinking" tokens after they already have enough information to answer, which wastes compute and is known as overthinking. The authors study this with a stepwise truncation protocol. They split the thinking trace into sentences, cut it off after each step, insert the </think> delimiter, and then examine the answer the model produces.

Two patterns appear. In length terms, a shorter thinking phase forces the model to write longer answers, but past a certain point the answer length stops changing while thinking keeps growing. In semantic terms, embeddings of the answers produced at each cutoff first wander widely and then settle into a stable region where they oscillate repetitively.

The authors call the boundary between these phases the Reasoning Completion Point (RCP). Because the RCP differs for each problem, they propose RCPD, a training-free early-exit method that watches how the rank of the termination token evolves during decoding. On AIME and GPQA with Qwen3 and DeepSeek-R1 models, RCPD cuts token use by up to 44% without losing accuracy. This gives a cheap, principled way to make test-time scaling more efficient.

Key points

  • Two-stage view of reasoning. The paper describes "thinking-content compensation" (length dynamics) and "semantic path convergence" (semantic dynamics). Both mark a shift from an active reasoning phase to a converged phase at the RCP.
  • Semantic convergence measurement.
    • For each truncation step, 32 answer continuations are sampled and embedded with the Qwen3 Embedding model.
    • The embeddings are projected with PCA and modeled as a Gaussian.
    • Convergence is tracked as the KL divergence to a terminal reference distribution, which falls and then plateaus.
    • Steps are defined as sentences segmented with NLTK.
  • How RCPD works. It monitors the rank of the end-of-thinking token (e.g., </think>), so it needs no extra model, classifier, or training and adds no computational overhead.
  • Evaluation setup. Experiments use Qwen3-8B, Qwen3-14B, Qwen3-30B-A3B, and DeepSeek-R1-Distill-8B on AIME (AIME24/25) and GPQA. Baselines are Full decoding, No-Think, BF, and DEER.
  • Results and limitations. RCPD saves up to 44% of tokens while preserving accuracy. The analysis does not cover closed-source APIs, tool-augmented agents, or multimodal models.

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

🎓University of Chinese Academy of SciencesBeijing · China147 works · 8,480 citations🔬Institute of Computing TechnologyBeijing · China51 works · 3,411 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 Zihao Wei, Liang Pang, Jiahao Liu, Wenjie Shi, Jingcheng Deng, Shicheng Xu, Zenghao Duan, Jingang Wang, Fei Sun, Huawei Shen and Xueqi Cheng. 2026. The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis. 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.1239

@inproceedings{wei2026evolution,
  title = {{The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis}},
  author = {Zihao Wei, Liang Pang, Jiahao Liu, Wenjie Shi, Jingcheng Deng, Shicheng Xu, Zenghao Duan, Jingang Wang, Fei Sun, Huawei Shen and Xueqi Cheng},
  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.1239},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1239}
}

References

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

  1. Pranjal Aggarwal and Sean Welleck. 2025. L1: controlling how long A reasoning model thinks with reinforcement learning. CoRR, abs/2503.04697.
  2. Steven Bird and Edward Loper. 2004. NLTK: the natural language toolkit. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, Barcelona, Spain, July 21-26, 2004 -Poster and Demonstration. ACL.
  3. Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. 2024a. Alphamath almost zero: Process supervision without process. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024…
  4. Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. 2024b. Step-level value preference optimization for mathematical reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7889–7903, Miami, Florida, USA. Association for Computational Linguistics. find
  5. Jeffrey Cheng and Benjamin Van Durme. 2024. Compressed chain of thought: Efficient reasoning through dense representations. CoRR, abs/2412.13171. find
  6. Dongkyu Cho, Amy B. Z. Zhang, Bilel Fehri, Sheng Wang, Rumi Chunara, Rui Song, and Hengrui Cai. 2025. Correct reasoning paths visit shared decision pivots. CoRR, abs/2509.21549. — [arXiv:2509.21549](https://arxiv.org/abs/2509.21549) find
  7. Muzhi Dai, Chenxu Yang, and Qingyi Si. 2025. S-GRPO: early exit via reinforcement learning in reasoning models. CoRR, abs/2505.07686. — [arXiv:2505.07686](https://arxiv.org/abs/2505.07686) find
  8. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. CoRR, abs/2501.12948.
  9. Soumya Suvra Ghosal, Souradip Chakraborty, Avinash Reddy, Yifu Lu, Mengdi Wang, Dinesh Manocha, Furong Huang, Mohammad Ghavamzadeh, and Amrit Singh Bedi. 2025. Does thinking more always help? mirage of test-time scaling in reasoning models. Preprint, arXiv:2506.04210. — [arXiv:2506.04210](https://arxiv.org/abs/2506.04210) find
  10. Google DeepMind. 2025. Gemini 3 pro model card. Model card, Google DeepMind. find
  11. Jinyi Han, Ying Huang, Ying Liao, Zishang Jiang, Xikun Lu, Haiquan Zhao, Xinyi Wang, Guanghao Zhou, Sihang Jiang, Jiaqing Liang, Weikang Zhou, Zeye Sun, Fei Yu, and Yanghua Xiao. 2025. Your models have thought enough: Training large reasoning models to stop overthinking. CoRR, abs/2509.23392. — [arXiv:2509.23392](https://arxiv.org/abs/2509.23392) find
  12. Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2024. Token-budget-aware LLM reasoning. CoRR, abs/2412.18547.
  13. Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training large language models to reason in a continuous latent space. CoRR, abs/2412.06769. find
  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. CoRR, abs/2508.05337. — [arXiv:2508.05337](https://arxiv.org/abs/2508.05337) find
  15. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, and 1 others. 2024. Openai o1 system card. CoRR, abs/2412.16720.
  16. Yu Kang, Xianghui Sun, Liangyu Chen, and Wei Zou. 2025. C3ot: Generating shorter chain-of-thought without compromising effectiveness. In AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA, USA, pages 24312–24320. AAAI P… find
  17. Kimi Team, Angang Du, Bofei Gao, Bowei Xing, and 1 others. 2025. Kimi k1.5: Scaling reinforcement learning with llms. CoRR, abs/2501.12599.
  18. Ayeong Lee, Ethan Che, and Tianyi Peng. 2025. How well do llms compress their own chain-of-thought? A token complexity approach. CoRR, abs/2503.01141. — [arXiv:2503.01141](https://arxiv.org/abs/2503.01141) find
  19. Melody Zixuan Li, Kumar Krishna Agrawal, Arna Ghosh, Komal Kumar Teru, Adam Santoro, Guillaume Lajoie, and Blake A. Richards. 2025. Tracing the representation geometry of language models from pretraining to post-training. CoRR, abs/2509.23024. — [arXiv:2509.23024](https://arxiv.org/abs/2509.23024) find
  20. Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s verify step by step. arXiv preprint arXiv:2305.20050.
  21. Xin Liu and Lu Wang. 2025. Answer convergence as a signal for early stopping in reasoning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 17907–17918, Suzhou, China. Association for Computational Linguistics.
  22. Haotian Luo, Li Shen, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Naiqiang Tan, Xiaochun Cao, and Dacheng Tao. 2025. O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning. CoRR, abs/2501.12570. find
  23. Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. 2025. Cot-valve: Length-compressible chain-of-thought tuning. CoRR, abs/2502.09601.
  24. MAA. 2025. Aime problems and solutions. find
  25. 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. CoRR, abs/2501.19393.
  26. 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. CoRR, abs/2602.09953.
  27. OpenAI, Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, and 1 others. 2025. gpt-oss-120b & gpt-oss-20b model card. Preprint, arXiv:2508.10925.
  28. Keqin Peng, Liang Ding, Yuanxin Ouyang, Meng Fang, and Dacheng Tao. 2025. Revisiting overthinking in long chain-of-thought from the perspective of self-doubt. CoRR, abs/2505.23480. — [arXiv:2505.23480](https://arxiv.org/abs/2505.23480) find
  29. Xiao Pu, Michael Saxon, Wenyue Hua, and William Yang Wang. 2025. THOUGHTTERMI-NATOR: benchmarking, calibrating, and mitigating overthinking in reasoning models. CoRR, abs/2504.13367. — [arXiv:2504.13367](https://arxiv.org/abs/2504.13367) find
  30. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. 2023. Gpqa: A graduate-level google-proof q&a benchmark. Preprint, arXiv:2311.12022.
  31. Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. 2025. CODI: compressing chain-of-thought into continuous space via self-distillation. CoRR, abs/2502.21074.
  32. Jinyan Su, Jennifer Healey, Preslav Nakov, and Claire Cardie. 2025. Between underthinking and overthinking: An empirical study of reasoning length and correctness in llms. CoRR, abs/2505.00127. — [arXiv:2505.00127](https://arxiv.org/abs/2505.00127) find
  33. Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Hanjie Chen, and Xia Ben Hu. 2025. Stop overthinking: A survey on efficient reasoning for large language models. CoRR, abs/2503.16419.
  34. Renliang Sun, Wei Cheng, Dawei Li, Haifeng Chen, and Wei Wang. 2025. Stop when enough: Adaptive early-stopping for chain-of-thought reasoning. CoRR, abs/2510.10103.
  35. Meituan LongCat Team, Anchun Gui, Bei Li, Bingyang Tao, Bole Zhou, Borun Chen, Chao Zhang, Chengcheng Han, Chenhui Yang, Chi Zhang, Chong Peng, Chuyu Zhang, Cong Chen, Fengcun Li, Gang Xu, Guoyuan Lin, Hao Jiang, Hao Liang, Haomin Fu, and 80 others. 2025. Longcat-flash-thinking technical report. CoR… — [arXiv:2509.18883](https://arxiv.org/abs/2509.18883) find
  36. Rui Wang, Hongru Wang, Boyang Xue, Jianhui Pang, Shudong Liu, Yi Chen, Jiahao Qiu, Derek Fai Wong, Heng Ji, and Kam-Fai Wong. 2025. Harnessing the reasoning economy: A survey of efficient reasoning for large language models. CoRR, abs/2503.24377.
  37. Yuyang Wu, Yifei Wang, Tianqi Du, Stefanie Jegelka, and Yisen Wang. 2025a. When more is less: Understanding chain-of-thought length in llms. CoRR, abs/2502.07266. find
  38. Yuyang Wu, Yifei Wang, Ziyu Ye, Tianqi Du, Stefanie Jegelka, and Yisen Wang. 2025b. When more is less: Understanding chain-of-thought length in llms. Preprint, arXiv:2502.07266. find
  39. Heming Xia, Yongqi Li, Chak Tou Leong, Wenjie Wang, and Wenjie Li. 2025. Tokenskip: Controllable chain-of-thought compression in llms. CoRR, abs/2502.12067.
  40. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, and 1 others. 2025a. Qwen3 technical report. Preprint, arXiv:2505.09388.
  41. Chenxu Yang, Qingyi Si, Yongjie Duan, Zheliang Zhu, Chenyu Zhu, Zheng Lin, Li Cao, and Weiping Wang. 2025b. Dynamic early exit in reasoning models. CoRR, abs/2504.15895. find
  42. Ping Yu, Jing Xu, Jason Weston, and Ilia Kulikov. 2024. Distilling system 2 into system 1. CoRR, abs/2407.06023. find
  43. Wenyuan Zhang, Shuaiyi Nie, Xinghua Zhang, Zefeng Zhang, and Tingwen Liu. 2025a. S1-bench: A simple benchmark for evaluating system 1 thinking capability of large reasoning models. CoRR, abs/2504.10368. — [arXiv:2504.10368](https://arxiv.org/abs/2504.10368) find
  44. Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025b. Qwen3 embedding: Advancing text embedding and reranking through foundation models. CoRR, abs/2506.05176.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wei-2026-evolution-thought-tracking-llm-overthinking/