Elamur Research⌕
☰
home / papers / 2026 / Solve-Detect-Verify: Inference-Time Scaling with Flexible Generative Verifier

Solve-Detect-Verify: Inference-Time Scaling with Flexible Generative Verifier

Paper · 2026 · Jianyuan Zhong, Zeju Li, Zhijian Xu et al.

Summary

Step-by-step reasoning in LLMs creates a cost problem. Solvers keep generating text after they already have an answer, and checking their work is expensive. Process-based verifiers such as GenPRM re-read the context at every step. Cheap "NoThinking"-style checks are fast but flag many false positives.

The authors propose two things:

  • FlexiVe, a generative verifier that reads the whole trace at once. It outputs a written critique and the index of the first wrong step. It has three modes: a short Fast mode, a full-reasoning Slow mode, and a Flex mode. Flex starts with several parallel fast checks and moves to slow reasoning only when the fast checks show uncertainty. Only the fast mode is trained, using GRPO with a reward for locating the first error correctly and a penalty on output length. This training also improves the slow mode.
  • Solve-Detect-Verify (SDV), an inference pipeline. It uses likelihood-based probing to notice when the solver has reached a complete answer and stops generation there. FlexiVe's feedback then guides a revision.

FlexiVe sets a new open-source best on ProcessBench, and SDV reaches 83.3% on AIME 2024. This matters because accuracy improves while compute is spent only where it is needed.

Key points

  • ProcessBench: FlexiVe beats GenPRM-32B while using about 2.3x fewer TFLOPS and about 15x less training data.
  • AIME 2024: SDV scores 83.3%, compared with 63.3% for a GenPRM Best-of-N setup, while using roughly one sixth of the verification tokens.
  • Effect of fast-mode GRPO training (average F1):
    • Think@1 rises from 70.9 to 75.6.
    • NoThink@4 rises from 37.1 to 58.6.
    • Flex@4 rises from 60.0 to 74.0.
  • Flex trade-off and speed:
    • With threshold τ = 0.8, Flex reaches 72.9 F1 using 3.0M tokens, versus 84.9 F1 and 23.1M tokens for Think@8.
    • Median latency is about 2 s, versus about 50 s for sequential GenPRM.
  • Transfer and limitations:
    • The Detect stage saves tokens on GPQA-Diamond, with a 3.53-point accuracy drop.
    • On OpenBookQA and ARC-Challenge it keeps or slightly improves accuracy.
    • The refinement loop fixes 4 of 16 initially wrong AIME 2024 answers.
    • Full-pipeline validation is mostly limited to math.
    • Early stopping depends on solvers whose expressed hesitation is well calibrated; this assumption is weaker for purely SFT-trained 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

🎓Chinese University of Hong KongHong Kong · China126 works · 14,073 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

Cite

Cite as Jianyuan Zhong, Zeju Li, Zhijian Xu, Xiangyu Wen, Kezhi Li and Qiang Xu. 2026. Solve-Detect-Verify: Inference-Time Scaling with Flexible Generative Verifier. 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.2190

@inproceedings{zhong2026solve,
  title = {{Solve-Detect-Verify: Inference-Time Scaling with Flexible Generative Verifier}},
  author = {Jianyuan Zhong, Zeju Li, Zhijian Xu, Xiangyu Wen, Kezhi Li and Qiang Xu},
  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.2190},
  url = {https://doi.org/10.18653/v1/2026.acl-long.2190}
}

References

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

  1. Afra Feyza Akyurek, Ekin Akyurek, Ashwin Kalyan, Peter Clark, Derry Tanti Wijaya, and Niket Tandon. 2023. RL4F: Generating natural language feedback with reinforcement learning for repairing model outputs. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Vo…
  2. Lukas Biewald. 2020. Experiment tracking with weights and biases. https://wandb.ai. Software available from wandb.com. ↗
  3. Jiefeng Chen, Jie Ren, Xinyun Chen, Chengrun Yang, Ruoxi Sun, and Sercan Ö. Arık. 2025. SETS: Leveraging self-verification and self-correction for improved test-time scaling. arXiv preprint arXiv:2501.19306. — [arXiv:2501.19306](https://arxiv.org/abs/2501.19306) find
  4. Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2024. Do not think that much for 2+3=? on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187.
  5. Alex Graves. 2016. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983. find
  6. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR).
  7. Hugging Face. 2024. Aime 2024 dataset. https://huggingface.co/datasets/HuggingFaceH4/ aime_2024. Accessed via Hugging Face Datasets. ↗
  8. Yuki Ichihara, Yuu Jinnai, Tetsuro Morimura, Kaito Ariu, Kenshi Abe, Mitsuki Sakamoto, and Eiji Uchibe. 2025. Evaluation of best-of-n sampling strategies for language model alignment. Preprint, arXiv:2502.12668. — [arXiv:2502.12668](https://arxiv.org/abs/2502.12668) find
  9. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, and 1 others. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221. find
  10. OpenCompass. 2025. Aime 2025 dataset. https://huggingface.co/datasets/opencompass/ AIME2025. Accessed via Hugging Face Datasets. ↗
  11. Daniel Kahneman. 2011. Thinking, fast and slow. Farrar, Straus and Giroux. find
  12. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024a. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300.
  13. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916.
  14. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Lee, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles (SOSP ’…
  15. Zhong-Zhi Li, Haotian Wang, Kaiyan Zhang, Yancheng He, Yujia Xie, Yuxiang Huang, Zhengliang Shi, HongCheng Li, Wenxuan Wang, Zhiwei He, Dian Yu, Haitao Mi, Dong Yu, Jie Tang, and AnBo Zhang. 2025. From system 1 to system 2: A survey of reasoning large language models. arXiv preprint arXiv:2502.17419…
  16. Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cognome. 2023. Let’s verify step by step. arXiv preprint arXiv:2305.20050.
  17. Stephanie C. Lin, Jacob Hilton, and Owain Evans. 2022. Teaching models to express their uncertainty in words. find
  18. Trans. Mach. Learn. Res., 2022. find
  19. Junnan Liu, Hongwei Liu, Linchen Xiao, Ziyi Wang, Kuikun Liu, Songyang Gao, Wenwei Zhang, Songyang Zhang, and Kai Chen. 2024. Are your llms capable of stable reasoning? arXiv preprint arXiv:2412.13147. find
  20. Zijun Liu, Peiyi Wang, Runxin Xu, Shirong Ma, Chong Ruan, Peng Li, Yang Liu, and Yu Wu. 2025. Inference-time scaling for generalist reward modeling. Preprint, arXiv:2504.02495.
  21. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR). find
  22. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y.K. Li, Y. Wu, and Daya Guo. 2024b. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300.
  23. Nishad Singhi, Hritik Bansal, Arian Hosseini, Aditya Grover, Kai-Wei Chang, Marcus Rohrbach, and Anna Rohrbach. 2025. When to solve, when to verify: Compute-optimal problem solving and generative verification for llm reasoning. Preprint, arXiv:2504.01005.
  24. Hanshi Sun, Momin Haider, Ruiqi Zhang, and 1 others. 2024. Fast best-of-n decoding via speculative rejection. In Advances in Neural Information Processing Systems (NeurIPS). find
  25. Gladys Tyen, Hassan Mansoor, Victor Carbune, Peter Chen, and Tony Mak. 2024. LLMs cannot find reasoning errors, but can correct them given the error location. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13894–13908, Bangkok, Thailand. Association for Computational L…
  26. Leandro von Werra, Lewis Schmid, Thomas Wolf, and Lewis Tunstall. 2020-2024. Trl: Transformer reinforcement learning. https://github.com/ huggingface/trl. ↗
  27. Jingxuan Wang, Yiming Ming, Zhengliang Shi, and 1 others. 2025. Inference-time scaling for complex tasks: Where we stand and what lies ahead. arXiv preprint arXiv:2504.00294. find
  28. Ke Wang, Houxing Ren, Aojun Zhou, and 1 others. 2024a. Mathcoder: Seamless code integration in llms for enhanced mathematical reasoning. In International Conference on Learning Representations (ICLR).
  29. Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. 2025. Reasoning models can be effective without thinking. Preprint, arXiv:2504.09858. find
  30. Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, and 1 others. 2024b. Math-Shepherd: Verify and reinforce LLMs step-by-step without human annotations. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL).
  31. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, and 1 others. 2023. Self-refine: Iterative refinement with self-feedback. In Advances in Neural Information Processing Systems (NeurIPS), volume 36. find
  32. 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 (ICLR).
  33. 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. In Advances in Neural Information Processing Systems, volume 35, pages 24824–24837. Curran Associates, Inc.
  34. Sean Welleck, Amanda Bertsch, Matthew Finlayson, and 1 others. 2024. From decoding to meta-generation: Inference-time algorithms for large language models.
  35. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, and 3 others. 2020. Transf…
  36. Qi, Xiu Li, and Bowen Zhou. 2025b. Genprm: Scaling test-time compute of process reward models via generative reasoning. Preprint, arXiv:2504.00891. — [arXiv:2504.00891](https://arxiv.org/abs/2504.00891) find
  37. Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2024a. Processbench: Identifying process errors in mathematical reasoning. Preprint, arXiv:2412.06559.
  38. Lianmin Zheng, Siyuan Zhuang, Zhuohan Li, Cody Hao Yu, Lequn Li, Haotian Chen, Joseph E. Gonzalez, Ion Stoica, and Jonathan Ragan-Kelley. 2024b. SGLang: Efficient and expressive structured generation for large language models. In Proceedings of the 18th Conference of the European Chapter of the Asso… find
  39. Jianyuan Zhong, Zeju Li, Zhijian Xu, Xiangyu Wen, and Qiang Xu. 2025. Dyve: Thinking fast and slow for dynamic process verification. Preprint, arXiv:2502.11157. find
  40. Yiming Wu, Zihan Sun, Sida Li, and 1 others. 2025a. Inference scaling laws: An empirical analysis of compute-optimal inference for llm problem-solving. In International Conference on Learning Representations (ICLR). find
  41. Aojun Zhou, Ke Wang, Zimu Lu, and 1 others. 2024a. Solving challenging math word problems using gpt-4 code interpreter with code-based self-verification. In International Conference on Learning Representations (ICLR).
  42. Yiming Wu, Zihan Sun, Sida Li, and 1 others. 2025b. Scaling flaws of verifier-guided search in mathematical reasoning. arXiv preprint arXiv:2502.00271. find
  43. Jin Peng Zhou, Charles Staats, Wenda Li, and 1 others. 2024b. Don’t trust: Verify-grounding llm quantitative reasoning with autoformalization. arXiv preprint arXiv:2403.18120. find
  44. Noah Xie, AI AUTOdidax, M Sarmad Parvez, Michael Song, Zhenqiao Zhang, Ziyu Chen, Shrimai Joshi, Robert Gmyr, Yufan Li, Siyuan Li, and 1 others. 2023. Reflexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems (NeurIPS), volume 36.
  45. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Sha, Thomas L Chen, Boyuan Rius, Yuxuan Du, Yang Liu, Zipeng Jiang, Tushar Han, and 1 others. 2023. Tree of thoughts: Deliberate problem solving with large language models. In Advances in Neural Information Processing Systems (NeurIPS), volume 36. find
  46. Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. 2022. Star: Bootstrapping reasoning with reasoning. In Advances in Neural Information Processing Systems (NeurIPS).
  47. Lunjun Zhang, Arian Hosseini, Hritik Bansal, Mehran Kazemi, Aviral Kumar, and Rishabh Agarwal. 2025. Generative verifiers: Reward modeling as next-token prediction. Preprint, arXiv:2408.15240.
  48. Eric Zhao, Pranjal Awasthi, and Sreenivas Gollapudi. 2025a. Sample, scrutinize and scale: Effective inference-time search by scaling verification. arXiv preprint arXiv:2502.00891. — [arXiv:2502.00891](https://arxiv.org/abs/2502.00891) find
  49. Jian Zhao, Runze Liu, Kaiyan Zhang, Zhimu Zhou, Junqi Gao, Dong Li, Jiafei Lyu, Zhouyi Qian, Biqing find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhong-2026-solve-detect-verify-inference-time/