Elamur Research⌕
☰
home / papers / 2026 / TrigReason: Trigger-Based Collaboration between Small and Large Reasoning Models

TrigReason: Trigger-Based Collaboration between Small and Large Reasoning Models

Paper · 2026 · Yi Zhao, Yajuan Peng, Cam-Tu Nguyen et al.

Summary

Large Reasoning Models (LRMs) solve hard problems well, but their long chains of thought make inference slow. Earlier work such as SpecReason speeds this up by having a Small Reasoning Model (SRM) draft each step and the LRM approve or reject it. The authors show this LRM-as-a-Judge setup is unreliable and wasteful, because the large model ends up rewriting much of the output. They identify three ways SRMs typically fail: poor initial planning (path divergence), inability to handle very hard steps (cognitive overload), and weak self-correction (recovery inability).

Their framework, TrigReason, lets the SRM do most of the reasoning. The LRM steps in only when one of three events occurs:

  • at the very start, to set a strategy;
  • when the SRM's confidence becomes abnormally high;
  • when the SRM falls into an unproductive loop.

On AIME24, AIME25 and GPQA-D, across several SRM–LRM pairs, accuracy stays on par with the full LRM and with SpecReason, while far more work shifts to the small model. This makes edge–cloud deployment cheaper and faster.

Key points

  • Judge unreliability: Four LRMs scored identical trajectories anywhere from 1.87 to 8.93. QwQ-32B rejected 50.1%–80.9% of correct SRM steps and up to 63.7% of its own steps.
  • Work offloaded to the SRM: Compared with SpecReason, TrigReason uses 1.70×–4.79× more SRM-generated tokens. On AIME24 with DeepSeek-R1-1.5B + QwQ-32B, the SRM's token share rises from 35.55% to 61.37%.
  • Edge–cloud gains: Relative to SpecReason, latency drops by 43.9% and API cost by 73.3%. SpecReason itself was 22.44% slower and 42.31% costlier than running the LRM alone.
  • Models and settings: SRMs were DeepSeek-R1-1.5B and Qwen3-0.6B; LRMs were QwQ-32B and Qwen3-30B-A3B-Thinking. Default settings were overload threshold ρ = 0.85, n = 20 priming steps, and m = 1 rectification step.
  • Other domains: With no domain-specific tuning, TrigReason scored 0.687 on BBH, above both SpecReason (0.663) and the LRM (0.675). On ARC it scored 0.948, above SpecReason (0.942) but slightly below the LRM (0.957). Code is available at github.com/QQQ-yi/TrigReason.

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

🎓Fudan UniversityShanghai · China202 works · 22,509 citations🎓Shanghai Jiao Tong UniversityShanghai · China212 works · 19,544 citations🎓Nanjing UniversityNanjing · China99 works · 8,510 citations🎓Wuhan UniversityWuhan · China47 works · 1,246 citations🎓University of GöttingenGöttingen · Germany20 works · 158 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 1Research subjects and languages the work is about.Reasoning · 1,685 works

Cite

Cite as Yi Zhao, Yajuan Peng, Cam-Tu Nguyen, Zuchao Li, Xiaoliang Wang, Xiaoming Fu and Hai Zhao. 2026. TrigReason: Trigger-Based Collaboration between Small and Large Reasoning Models. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.333

@inproceedings{zhao2026trigreason,
  title = {{TrigReason: Trigger-Based Collaboration between Small and Large Reasoning Models}},
  author = {Yi Zhao, Yajuan Peng, Cam-Tu Nguyen, Zuchao Li, Xiaoliang Wang, Xiaoming Fu and Hai Zhao},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.333},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.333}
}

References

27 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 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.
  2. Wasi Uddin Ahmad, Sean Narenthiran, Somshubra Majumdar, Aleksander Ficek, Siddhartha Jain, Jocelyn Huang, Vahid Noroozi, and Boris Ginsburg. 2025. Opencodereasoning: Advancing data distillation for competitive coding. Preprint, arXiv:2504.01943.
  3. AIME. 2024. Aime 2024 dataset. https://huggingface.co/datasets/HuggingFaceH4/ aime_2024. ↗
  4. AIME. 2025. Aime 2025 dataset. https://huggingface.co/datasets/math-ai/aime25. ↗
  5. Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv:1803.05457v1.
  6. Quy-Anh Dang and Chris Ngo. 2025. Reinforcement learning for reasoning in small llms: What works and what doesn’t. Preprint, arXiv:2503.16219. find
  7. DeepSeek-AI. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint, arXiv:2501.12948.
  8. DeepSeek API. 2025. DeepSeek API Documentation. https://api-docs.deepseek.com/. Accessed: 2025-09-20. ↗
  9. Yu Kang, Xianghui Sun, Liangyu Chen, and Wei Zou. 2024. C3ot: Generating shorter chain-of-thought without compromising effectiveness. Preprint, arXiv:2412.11664.
  10. Qwen Team. 2025b. Qwq-32b: Embracing the power of reinforcement learning. https://qwenlm.github.io/blog/qwq-32b/. ↗
  11. Mirac Suzgun, Nathan Scales, Nathanael Schärli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc Le, Ed Chi, Denny Zhou, and Jason Wei. 2023. Challenging BIG-bench tasks and whether chain-of-thought can solve them. In Findings of the Association for Computational Linguistics: A…
  12. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837.
  13. Yifan Wu, Jingze Shi, Bingheng Wu, Jiayi Zhang, Xiaotian Lin, Nan Tang, and Yuyu Luo. 2025. Concise reasoning, big gains: Pruning long reasoning trace with difficulty-aware prompting. Preprint, arXiv:2505.19716. — [arXiv:2505.19716](https://arxiv.org/abs/2505.19716) find
  14. Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, and Wenjie Li. 2025. Tokenskip: Controllable chain-of-thought compression in llms. Preprint, arXiv:2502.12067.
  15. Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025. Chain of draft: Thinking faster by writing less. Preprint, arXiv:2502.18600.
  16. Kimi Team. 2025. Kimi k2: Open agentic intelligence. Preprint, arXiv:2507.20534.
  17. Junjie Yang, Ke Lin, and Xing Yu. 2025. Think when you need: Self-adaptive chain-of-thought learning. Preprint, arXiv:2504.03234. — [arXiv:2504.03234](https://arxiv.org/abs/2504.03234) find
  18. 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. Preprint, arXiv:2501.12570. find
  19. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2024a. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36. find
  20. Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. 2025. Cot-valve: Length-compressible chain-of-thought tuning. Preprint, arXiv:2502.09601.
  21. Yao Yao, Zuchao Li, and Hai Zhao. 2024b. GoT: Effective graph-of-thought reasoning in language models. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 2901–2921, Mexico City, Mexico. Association for Computational Linguistics. find
  22. OpenAI. 2024. Introducing openai o1-preview. https://openai.com/index/ introducing-openai-o1-preview/. ↗
  23. Rui Pan, Yinwei Dai, Zhihao Zhang, Gabriele Oliaro, Zhihao Jia, and Ravi Netravali. 2025. Specreason: Fast and accurate inference-time compute via speculative reasoning. Preprint, arXiv:2504.07891. — [arXiv:2504.07891](https://arxiv.org/abs/2504.07891) find
  24. Yi Zhao, Zuchao Li, and Hai Zhao. 2025a. Iam: Efficient inference through attention mapping between different-scale llms. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 19522– 19533. find
  25. Yi Zhao, Zuchao Li, Hai Zhao, Baoyuan Qi, and Liu Guoming. 2025b. Dac: A dynamic attention-aware approach for task-agnostic prompt compression. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 19395–19407. find
  26. Yi Zhao, Yajuan Peng, Cam-Tu Nguyen, Zuchao Li, Xiaoliang Wang, Hai Zhao, and Xiaoming Fu. 2025c. Smallkv: Small model assisted compensation of kv cache compression for efficient llm inference. arXiv preprint arXiv:2508.02751. — [arXiv:2508.02751](https://arxiv.org/abs/2508.02751) find
  27. Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Jeff Huang, Chuyue Sun, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E. Gonzalez, Clark Barrett, and Ying Sheng. 2023. Efficiently programming large language models using sglang. Preprint, arXiv:2312.07104.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhao-2026-trigreason-trigger-based-collaboration-between/