Elamur Research⌕
☰
home / papers / 2026 / EpiCaR: Knowing What You Don't Know Matters for Better Reasoning in LLMs

EpiCaR: Knowing What You Don't Know Matters for Better Reasoning in LLMs

Paper · 2026 · Jewon Yeom, Jaewon Sok, Seonghyeon Park et al.

Summary

Iterative self-training methods such as STaR and ReST improve LLM reasoning by fine-tuning only on the model's own correct solutions. The authors argue this has a hidden price. Because the model never sees where good reasoning ends and flawed reasoning begins, it becomes overconfident and loses its sense of uncertainty, a pattern they connect to model collapse. EpiCaR treats reasoning training as an epistemic problem: the model should learn both to solve tasks and to judge whether its own solutions deserve trust. Within an iterative SFT loop, each model is trained jointly on reasoning and on explicit yes/no self-evaluation of its answers. Confidence is read from the probability of the "yes" token.

On Llama-3 and Qwen-3 models trained on MATH, EpiCaR improves accuracy and calibration together, measured by ECE, Brier Score and AUROC. The gains also carry over to GSM8K and MBPP. The work matters because a single model's internal confidence can replace expensive external verifiers and reduce the number of samples needed at inference time.

Key points

  • Evaluated on Llama-3 and Qwen-3 families up to 8B parameters. Gains are clearest at about 3B and above, while Llama-3-1B produces too few correct paths to learn useful self-evaluation.
  • Using confidence-weighted voting (CISC), EpiCaR with K = 10 samples matches STaR's K = 30 majority vote, which cuts total inference compute roughly 3×.
  • As a single-model approach, it avoids the roughly 21× inference overhead of V-STaR's separate verifier at K = 10.
  • Out-of-distribution, ranking quality (AUROC) stays strong on GSM8K, but absolute calibration (raw ECE) is less stable than on in-distribution MATH.
  • Training on self-evaluation lowers sensitivity to prompt templates, with up to 44% less cross-template variance than zero-shot baselines such as STaR.

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

🎓Seoul National UniversitySeoul · South Korea43 works · 45,969 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 Jewon Yeom, Jaewon Sok, Seonghyeon Park, Jeongjae Park and Taesup Kim. 2026. EpiCaR: Knowing What You Don't Know Matters for Better Reasoning in LLMs. 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.1026

@inproceedings{yeom2026epicar,
  title = {{EpiCaR: Knowing What You Don't Know Matters for Better Reasoning in LLMs}},
  author = {Jewon Yeom, Jaewon Sok, Seonghyeon Park, Jeongjae Park and Taesup Kim},
  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.1026},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1026}
}

References

38 references, 19 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and 1 others. 2021. Program synthesis with large language models. arXiv preprint arXiv:2108.07732.
  2. Glenn W Brier. 1950. Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1):1–3. find
  3. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
  4. Mehul Damani, Isha Puri, Stewart Slocum, Idan Shenfeld, Leshem Choshen, Yoon Kim, and Jacob Andreas. 2025. Beyond binary rewards: Training lms to reason about their uncertainty. arXiv preprint arXiv:2507.16806. find
  5. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  6. Abhimanyu Dubey, Akhil Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy — The Llama 3 Herd of Models https://deepmind.google/technologies/gemini/Yang, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. ↗
  7. Tingchen Fu, Deng Cai, Lemao Liu, Shuming Shi, and Rui Yan. 2024. Disperse-then-merge: Pushing the limits of instruction tuning via alignment tax reduction. In Findings of the Association for Computational Linguistics: ACL 2024, pages 2967–2985. Association for Computational Linguistics. find
  8. Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alex Ahern, Miaosen Wang, Chenjie Gu, and 1 others. 2023. Reinforced self-training (rest) for language modeling. arXiv preprint arXiv:2308.08998.
  9. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International Conference on Machine Learning, pages 1321–1330. PMLR. find
  10. 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. NeurIPS.
  11. Arian Hosseini, Xingdi Yuan, Nikolay Malkin, Aaron Courville, Alessandro Sordoni, and Rishabh Agarwal. 2024. V-STaR: Training verifiers for self-taught reasoners. In First Conference on Language Modeling (COLM).
  12. Tiancheng Hu, Benjamin Minixhofer, and Nigel Collier. 2025. Navigating the alignment-calibration trade-off: A pareto-superior frontier via model merging. arXiv preprint arXiv:2510.17426. find
  13. Eyke Hüllermeier and Willem Waegeman. 2021. Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods. Machine Learning, 110:457–506. find
  14. Hyunbin Jin, Je Won Yeom, Seunghyun Bae, and Taesup Kim. 2025. "well, keep thinking": Enhancing llm reasoning with adaptive injection decoding. arXiv preprint arXiv:2503.10167. find
  15. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zachary Dodds, Nova DasSarma, Eli Tran-Johnson, and 1 others. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221. find
  16. Sanyam Kapoor, Manley Roberts, Katherine Collins, Arka Pal, Nate Gruver, Umang Bhatt, Adrian Weller, Samuel Dooley, Micah Goldblum, and Andrew Gordon Wilson. 2024. Large language models must be taught to know what they don’t know. In Advances in Neural Information Processing Systems, volume 36. find
  17. Yoonjeon Kim, Doohyuk Jang, and Eunho Yang. 2025. Meta-awareness enhances reasoning models: Self-alignment reinforcement learning. arXiv preprint arXiv:2510.03259. — [arXiv:2510.03259](https://arxiv.org/abs/2510.03259) find
  18. Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations (ICLR).
  19. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. 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 (SOSP), pages 61…
  20. Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. Teaching models to express their uncertainty in words. Transactions on Machine Learning Research. find
  21. Yong Lin, Hangyu Lin, Wei Xiong, Shizhe Diao, Jianmeng Liu, Jipeng Zhang, Rui Pan, Haoxiang Wang, Wenbin Hu, Hanning Zhang, Hanze Dong, Renjie Pi, Han Zhao, Nan Jiang, Heng Ji, Yuan Yao, and Tong Zhang. 2024. Mitigating the alignment tax of rlhf. In Proceedings of the 2024 Conference on Empirical Me…
  22. Xin Liu, Muhammad Khalifa, and Lu Wang. 2024. Lit-Cab: Lightweight language model calibration over short- and long-form responses. In The Twelfth International Conference on Learning Representations. find
  23. Keming Lu, Bowen Yu, Fei Huang, Yang Fan, Runji Lin, and Chang Zhou. 2024. Online merging optimizers for boosting rewards and mitigating tax in alignment. arXiv preprint arXiv:2405.17931. — [arXiv:2405.17931](https://arxiv.org/abs/2405.17931) find
  24. OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.
  25. Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems (NeurIPS).
  26. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. find
  27. Maohao Shen, Subhro Das, Kristjan Greenewald, Prasanna Sattigeri, Gregory Wornell, and Soumya Ghosh. 2024. Thermometer: Towards universal calibration for large language models. In Proceedings of the 41st International Conference on Machine Learning, pages 44607–44631. PMLR. find
  28. Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and Ross Anderson. 2023. The curse of recursion: Training on generated data makes models forget. arXiv preprint arXiv:2305.17493.
  29. Amir Taubenfeld, Tom Sheffer, Eran Ofek, Amir Feder, Ariel Goldstein, Zorik Gekhman, and Gal Yona. 2025. Confidence improves self-consistency in LLMs. In Findings of the Association for Computational Linguistics: ACL 2025, pages 20090–20111, Vienna, Austria. Association for Computational Linguistics… find
  30. Ante Wang, Linfeng Song, Ye Tian, Baolin Peng, Lifeng Jin, Haitao Mi, Jinsong Su, and Dong Yu. 2024. Self-consistency boosts calibration for math reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 6023–6029, Miami, Florida, USA. Association for Computational L… find
  31. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations.
  32. 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 (NeurIPS), volume 35, pages 24824–24837.
  33. Yuxi Xia, Pedro Henrique Luz De Araujo, Klim Zaporojets, and Benjamin Roth. 2025. Influences on LLM calibration: A study of response agreement, loss functions, and prompt styles. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages…
  34. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025a. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  35. Ruihan Yang, Caiqi Zhang, Zhisong Zhang, Xinting Huang, Sen Yang, Nigel Collier, Dong Yu, and Deqing Yang. 2025b. LoGU: Long-form generation with uncertainty expressions. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 18947–1…
  36. Dongkeun Yoon, Seungone Kim, Sohee Yang, Sunkyoung Kim, Soyeon Kim, Yongil Kim, Eunbi Choi, Yireun Kim, and Minjoon Seo. 2025. Reasoning models better express their confidence. In The 39th Conference on Neural Information Processing Systems (NeurIPS). find
  37. Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. 2022. Star: Bootstrapping reasoning with reasoning. In Advances in Neural Information Processing Systems, volume 35, pages 15476–15488.
  38. Anqi Zhang, Yulin Chen, Jane Pan, Chen Zhao, Aurojit Panda, Jinyang Li, and He He. 2025. Reasoning models know when they’re right: Probing hidden states for self-verification. In The 2nd Conference on Language Modeling (COLM). find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yeom-2026-epicar-knowing-what-you-don/