Elamur Research⌕
☰
home / papers / 2026 / RFS-Guard: Detecting Reasoning Hallucinations via Cross-Phase Routing Focus in Large Reasoning Models

RFS-Guard: Detecting Reasoning Hallucinations via Cross-Phase Routing Focus in Large Reasoning Models

Paper · 2026 · Zihang Liu, Fang Zhouhua, Hui Liu et al.

Summary

Large reasoning models (LRMs) can produce fluent chains of thought that still hide small arithmetic slips or broken constraints, which then lead to wrong final answers. Existing detectors often depend on external checking, repeated sampling, or token-level confidence, and these approaches suit long multi-step outputs poorly. The authors look instead at how information travels from the reasoning phase to the answer phase.

They define the Routing Focus Score (RFS), a step-level measure of how much cross-step attention concentrates on semantically similar steps. Semantic similarity comes from hidden-state cosine similarity. RFS-Guard combines this score with multi-hop backtracking of the reasoning flow. The result is a training-free detector that also localizes errors.

Their main empirical finding is that hallucinations tend to come with higher reasoning–answer RFS. The authors read this as "routing collapse": the model reinforces its own earlier claims instead of checking them. On MATH, Science and MultiHopQA, using R1-7B and Qwen3-8B, RFS-Guard mostly beats the baselines. This matters because it offers a cheap internal signal that needs no external tools, though it works only with open-weight models.

Key points

  • Empirical study setup: The pilot study pooled MATH500, AIME25 and minervamath and ran DeepSeek-R1-Distill-Qwen-7B and Qwen3-8B on them. Qwen3-235B judged each answer step five times. Model labels agreed with a human spot-check 86% of the time.
  • Error types: In a sample of 100 hallucinated steps, 92% were concrete arithmetic errors or explicit constraint violations.
  • Head and step pooling: Attention heads are chosen per sample by lowest entropy. Token attention is pooled to step level, and positional (recency) bias is corrected.
  • Localization results: Compared with baselines such as SelfCheckGPT, P(True), PPL, CCP, EigenScore and LengthScore, RFS-Guard mostly achieved the best results. Gains reached up to 8.0%, with a few small regressions.
  • Ablation and limits: Removing reasoning-flow backtracking lowered scores by 2.5% for R1-7B and 4.6% for Qwen3-8B on average. On long traces of about 100 steps, the drop reached roughly 5–6%. The method needs access to attention weights and hidden states, and the signal it gives is correlational rather than causal.

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

🎓Zhejiang UniversityHangzhou · China172 works · 12,988 citations🏢Ant GroupHangzhou · China42 works · 1,131 citations

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,426 works

Cite

Cite as Zihang Liu, Fang Zhouhua, Hui Liu, Zhiwei Liu, Yong Li and Haishuai Wang. 2026. RFS-Guard: Detecting Reasoning Hallucinations via Cross-Phase Routing Focus in Large Reasoning Models. 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.885

@inproceedings{liu2026rfs,
  title = {{RFS-Guard: Detecting Reasoning Hallucinations via Cross-Phase Routing Focus in Large Reasoning Models}},
  author = {Zihang Liu, Fang Zhouhua, Hui Liu, Zhiwei Liu, Yong Li and Haishuai Wang},
  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.885},
  url = {https://doi.org/10.18653/v1/2026.acl-long.885}
}

References

48 references, 18 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Samir Abdaljalil, Hasan Kurban, Parichit Sharma, Erchin Serpedin, and Rachad Atat. 2025. Sindex: Semantic inconsistency index for hallucination detection in llms. CoRR, abs/2503.05980. — [arXiv:2503.05980](https://arxiv.org/abs/2503.05980) find
  2. Amos Azaria and Tom M. Mitchell. 2023. The internal state of an LLM knows when it’s lying. In EMNLP (Findings), Findings of ACL, pages 967–976. Association for Computational Linguistics.
  3. Jakub Binkowski, Denis Janiak, Albert Sawczyn, Bogdan Gabrys, and Tomasz Kajdanowicz. 2025. Hallucination detection in llms using spectral features of attention maps. In EMNLP, pages 24354–24385. Association for Computational Linguistics.
  4. Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. 2024a. INSIDE: llms’ internal states retain the power of hallucination detection. In ICLR. OpenReview.net. find
  5. Jianghao Chen, Junhong Wu, Yangyifan Xu, and Jiajun Zhang. 2025. LADM: long-context training data selection with attention-based dependency measurement for llms. In ACL (1), pages 3076–3090. Association for Computational Linguistics.
  6. Shiqi Chen, Miao Xiong, Junteng Liu, Zhengxuan Wu, Teng Xiao, Siyang Gao, and Junxian He. 2024b. In-context sharpness as alerts: An inner representation perspective for hallucination mitigation. In ICML, Proceedings of Machine Learning Research, pages 7553–7567. PMLR / OpenReview.net. find
  7. DeepSeek-AI. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. CoRR, abs/2501.12948.
  8. Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2024. Chain-of-verification reduces hallucination in large language models. In ACL (Findings), Findings of ACL, pages 3563–3578. Association for Computational Linguistics.
  9. Ekaterina Fadeeva, Aleksandr Rubashevskii, Artem Shelmanov, Sergey Petrakov, Haonan Li, Hamdy Mubarak, Evgenii Tsymbalov, Gleb Kuzmin, Alexander Panchenko, Timothy Baldwin, Preslav Nakov, and Maxim Panov. 2024. Fact-checking the output of large language models via token-level uncertainty quantificat…
  10. Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. 2024. Detecting hallucinations in large language models using semantic entropy. Nat., 630(8017):625–630. find
  11. Junjie Hu, Gang Tu, Cheng Shengyu, JinXin Li, Jinting Wang, Rui Chen, Zhilong Zhou, and Dongbo Shan. 2026. HARP: Hallucination detection via reasoning subspace projection. In The Fourteenth International Conference on Learning Representations. find
  12. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2025. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Trans. Inf. Syst., 43(2):42:1– 42:55…
  13. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bowman, Stanislav Fort, and 17 others. 2022.… find
  14. Haoqiang Kang, Juntong Ni, and Huaxiu Yao. 2023. Ever: Mitigating hallucination in large language models through real-time verification and rectification. CoRR, abs/2311.09114. — [arXiv:2311.09114](https://arxiv.org/abs/2311.09114) find
  15. Haoxi Li, Sikai Bai, Jie Zhang, and Song Guo. 2025a. Core: Enhancing metacognition with label-free self-evaluation in lrms. CoRR, abs/2507.06087. — [arXiv:2507.06087](https://arxiv.org/abs/2507.06087) find
  16. Kenneth Li, Oam Patel, Fernanda B. Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. Inference-time intervention: Eliciting truthful answers from a language model. In NeurIPS.
  17. Yinghao Li, Rushi Qiang, Lama Moukheiber, and Chao Zhang. 2025b. Language model uncertainty quantification with attention chain. CoRR, abs/2503.19168. — [arXiv:2503.19168](https://arxiv.org/abs/2503.19168) find
  18. Zihang Liu, Jiawei Guo, Hao Zhang, Hongyang Chen, Jiajun Bu, and Haishuai Wang. 2025. Long-form hallucination detection with self-elicitation. In ACL (Findings), Findings of ACL, pages 4082–4100. Association for Computational Linguistics. find
  19. Haolang Lu, Yilian Liu, Jingxin Xu, Guoshun Nan, Yuanlong Yu, Zhican Chen, and Kun Wang. 2025. Auditing meta-cognitive hallucinations in reasoning large language models. CoRR, abs/2505.13143. — [arXiv:2505.13143](https://arxiv.org/abs/2505.13143) find
  20. Andrey Malinin and Mark J. F. Gales. 2021a. Uncertainty estimation in autoregressive structured prediction. In ICLR. OpenReview.net. find
  21. Andrey Malinin and Mark J. F. Gales. 2021b. Uncertainty estimation in autoregressive structured prediction. In ICLR. OpenReview.net. find
  22. Potsawee Manakul, Adian Liusie, and Mark J. F. Gales. 2023. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. In EMNLP, pages 9004–9017. Association for Computational Linguistics. find
  23. Ning Miao, Yee Whye Teh, and Tom Rainforth. 2024. Selfcheck: Using llms to zero-shot check their own step-by-step reasoning. In ICLR. OpenReview.net. find
  24. Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. Factscore: Fine-grained atomic evaluation of factual precision in long form text generation. In EMNLP, pages 12076–12100. Association for Computational Linguist…
  25. Niels Mündler, Jingxuan He, Slobodan Jenko, and Martin T. Vechev. 2024. Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation. In ICLR. OpenReview.net.
  26. Yifu Qiu, Varun R. Embar, Yizhe Zhang, Navdeep Jaitly, Shay B. Cohen, and Benjamin Han. 2025. Eliciting in-context retrieval and reasoning for long-context large language models. In ACL (Findings), Findings of ACL, pages 3176–3192. Association for Computational Linguistics. find
  27. Subhey Sadi Rahman, Md. Adnanul Islam, Md. Mahbub Alam, Musarrat Zeba, Md. Abdur Rahman, Sadia Sultana Chowa, Mohaimenul Azam Khan Raiaan, and Sami Azam. 2026. Hallucination to truth: a review of fact-checking and factuality evaluation in large language models. Artif. Intell. Rev., 59(2):70. find
  28. Jie Ren, Jiaming Luo, Yao Zhao, Kundan Krishna, Mohammad Saleh, Balaji Lakshminarayanan, and Peter J. Liu. 2023. Out-of-distribution detection and selective generation for conditional language models. In ICLR. OpenReview.net. find
  29. Gaurang Sriramanan, Siddhant Bharti, Vinu Sankar Sadasivan, Shoumik Saha, Priyatham Kattakinda, and Soheil Feizi. 2024. Llm-check: Investigating detection of hallucinations in large language models. In NeurIPS. find
  30. Zhongxiang Sun, Qipeng Wang, Haoyu Wang, Xiao Zhang, and Jun Xu. 2026. Mechanistic detection and mitigation of hallucination in large reasoning models. In The Fourteenth International Conference on Learning Representations. find
  31. Zhongxiang Sun, Xiaoxue Zang, Kai Zheng, Jun Xu, Xiao Zhang, Weijie Yu, Yang Song, and Han Li. 2025. Redeep: Detecting hallucination in retrieval-augmented generation via mechanistic interpretability. In ICLR. OpenReview.net. find
  32. Qwen Team. 2025. Qwen3 technical report. CoRR, abs/2505.09388.
  33. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NIPS, pages 5998–6008.
  34. Changyue Wang, Weihang Su, Qingyao Ai, and Yiqun Liu. 2026. Joint evaluation of answer and reasoning consistency for hallucination detection in large reasoning models. In AAAI, pages 33377–33385. AAAI Press. find
  35. Lei Wang. 2025. Seredeep: Hallucination detection in retrieval-augmented models via semantic entropy and context-parameter fusion. CoRR, abs/2505.07528. — [arXiv:2505.07528](https://arxiv.org/abs/2505.07528) find
  36. Yanbo Wang, Yongcan Yu, Jian Liang, and Ran He. 2025. A comprehensive survey on trustworthiness in reasoning with large language models. CoRR, abs/2509.03871. find
  37. Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Hu, Jie Huang, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, Cosmo Du, and Quoc V. Le. 2024. Long-form factuality in large language models. In NeurIPS.
  38. Wenhao Wu, Yizhong Wang, Guangxuan Xiao, Hao Peng, and Yao Fu. 2025. Retrieval head mechanistically explains long-context factuality. In ICLR. Open-Review.net.
  39. Zidi Xiong, Shan Chen, Zhenting Qi, and Himabindu Lakkaraju. 2025. Measuring the faithfulness of thinking drafts in large reasoning models. CoRR, abs/2505.13774. — [arXiv:2505.13774](https://arxiv.org/abs/2505.13774) find
  40. Zijun Yao, Yantao Liu, Yanxu Chen, Jianhui Chen, Junfeng Fang, Lei Hou, Juanzi Li, and Tat-Seng Chua. 2025. Are reasoning models more prone to hallucination? CoRR, abs/2505.23646. — [arXiv:2505.23646](https://arxiv.org/abs/2505.23646) find
  41. Xinyue Zeng, Junhong Lin, Yujun Yan, Feng Guo, Liang Shi, Jun Wu, and Dawei Zhou. 2026. Halluguard: Demystifying data-driven and reasoning-driven hallucinations in LLMs. In The Fourteenth International Conference on Learning Representations. find
  42. Zhiyuan Zeng, Qinyuan Cheng, Zhangyue Yin, Yunhua Zhou, and Xipeng Qiu. 2025. Revisiting the test-time scaling of o1-like models: Do they truly possess test-time scaling capabilities? In ACL (1), pages 4651– 4665. Association for Computational Linguistics.
  43. Anqi Zhang, Yulin Chen, Jane Pan, Chen Zhao, Aurojit Panda, Jinyang Li, and He He. 2025a. Reasoning models know when they’re right: Probing hidden states for self-verification. CoRR, abs/2504.05419. find
  44. Guibin Zhang, Yuxiang Zhang, Moayad Aloqaily, Haolang Lu, Kun Wang, Jing Liang, Xingjun Ma, Yu-Gang Jiang, and Qingsong Wen. 2026. Unraveling hallucination in large reasoning models: A topological perspective. find
  45. Muru Zhang, Ofir Press, William Merrill, Alisa Liu, and Noah A. Smith. 2024a. How language model hallucinations can snowball. In ICML, Proceedings of Machine Learning Research, pages 59670–59684. PMLR / OpenReview.net.
  46. Wenqi Zhang, Yongliang Shen, Linjuan Wu, Qiuying Peng, Jun Wang, Yueting Zhuang, and Weiming Lu. 2024b. Self-contrast: Better reflection through inconsistent solving perspectives. In ACL (1), pages 3602–3622. Association for Computational Linguistics.
  47. Zhenliang Zhang, Xinyu Hu, Huixuan Zhang, Junzhe Zhang, and Xiaojun Wan. 2025b. ICR probe: Tracking hidden state dynamics for reliable hallucination detection in llms. In ACL (1), pages 17986–18002. Association for Computational Linguistics.
  48. Xinran Zhao, Hongming Zhang, Xiaoman Pan, Wenlin Yao, Dong Yu, Tongshuang Wu, and Jianshu Chen. 2024. Fact-and-reflection (far) improves confidence calibration of large language models. In ACL (Findings), Findings of ACL, pages 8702–8718. Association for Computational Linguistics. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liu-2026-rfs-guard-detecting-reasoning-hallucinations/