Elamur Research⌕
☰
home / papers / 2026 / MTRouter: Cost-Aware Multi-Turn LLM Routing with History–Model Joint Embeddings

MTRouter: Cost-Aware Multi-Turn LLM Routing with History–Model Joint Embeddings

Paper · 2026 · Yiqun Zhang, Hao Li, Zihan Wang et al.

Summary

Agentic tasks with many turns often need dozens of LLM calls per episode, so costs add up quickly. Using one frontier model throughout is expensive, and picking a single model at the start (episode-level routing) ignores that turns vary in difficulty. MTRouter tackles turn-level routing: at each step it chooses a model from a fixed pool while respecting a per-episode cost budget and a turn cap. It embeds the interaction history with a frozen text encoder. Each candidate model is embedded from its metadata (pricing, context limit, knowledge cut-off) plus a learned residual vector. A small MLP takes the combined history–model vector and predicts the final episode outcome. This estimator is trained offline on logged trajectories, using terminal scores reduced by penalties for detected errors. On ScienceWorld, MTRouter beats GPT-5 at much lower cost. On HLE, it matches GPT-5's accuracy closely while cutting cost substantially, and the gains also hold on out-of-distribution tasks. The work shows that a lightweight learned router can replace heavier LLM-based routers in long-horizon agent workflows.

Key points

  • On ScienceWorld (test), the average score rises from 48.4 with GPT-5 to 53.8, and total cost drops by 58.7%. On HLE (test), accuracy stays competitive with GPT-5 while cost falls by 43.4%.
  • The training data are trajectories from random-router and single-model runs. History is truncated to 8192 tokens, keeping the task block and the newest context. Error penalties depend on error severity and on how far into the episode the error occurs. Episodes are capped at 50 steps for ScienceWorld and 30 for HLE.
  • After an error, MTRouter keeps the same model in about 90.2% of cases on ScienceWorld and 80.9% on HLE, compared with 38.3% and 66.4% for Router-R1. Despite switching less, MTRouter recovers more often on the next turn.
  • Models take on specialized roles. On HLE, DeepSeek is over-used for search (lift 1.66), GPT-5 for python (1.51), and Kimi for browse (1.98).
  • The router learns only offline and cannot adapt online. Code is available at github.com/ZhangYiqun018/MTRouter.

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

🔬Shanghai Artificial Intelligence LaboratoryShanghai · China152 works · 13,164 citations🎓Northeastern UniversityShenyang · China53 works · 2,205 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.Language models & pretraining · 3,894 works

Cite

Cite as Yiqun Zhang, Hao Li, Zihan Wang, Shi Feng, Xiaocui Yang, Daling Wang, Bo Zhang, Lei Bai and Shuyue Hu. 2026. MTRouter: Cost-Aware Multi-Turn LLM Routing with History–Model Joint Embeddings. 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.2045

@inproceedings{zhang2026mtrouter,
  title = {{MTRouter: Cost-Aware Multi-Turn LLM Routing with History–Model Joint Embeddings}},
  author = {Yiqun Zhang, Hao Li, Zihan Wang, Shi Feng, Xiaocui Yang, Daling Wang, Bo Zhang, Lei Bai and Shuyue Hu},
  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.2045},
  url = {https://doi.org/10.18653/v1/2026.acl-long.2045}
}

References

28 references, 15 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Lingjiao Chen, Matei Zaharia, and James Zou. 2023. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176.
  2. Shuhao Chen, Weisen Jiang, Baijiong Lin, James Kwok, and Yu Zhang. 2024. Routerdc: Query-based router by dual contrastive learning for assembling large language models. Advances in Neural Information Processing Systems, 37:66305–66328. find
  3. Tri Dao. 2023. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691.
  4. Pengfei Gao and Chao Peng. 2025. More with less: An empirical study of turn-control strategies for efficient coding agents. Preprint, arXiv:2510.16786. — [arXiv:2510.16786](https://arxiv.org/abs/2510.16786) find
  5. Shuyue Hu, Siyue Ren, Yang Chen, Chunjiang Mu, Jinyi Liu, Zhiyao Cui, Yiqun Zhang, Hao Li, Dongzhan Zhou, Jia Xu, and 1 others. 2025. Hands-on llm-based agents: A tutorial for general audiences. find
  6. Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. 2024. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974.
  7. Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. 2023. Swe-bench: Can language models resolve real-world github issues? arXiv preprint arXiv:2310.06770.
  8. Logan Kilpatrick and Zach Gleicher. 2025. Gemini 2.5 Flash-Lite is now stable and generally available. Official Google Developers blog. Accessed: 2026-04-18. find
  9. Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, and 1 others. 2025. Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556.
  10. Grégoire Mialon, Clémentine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. 2023. Gaia: a benchmark for general ai assistants. Preprint, arXiv:2311.12983.
  11. MiniMax. 2025. Minimax m2 & agent: Ingenious in simplicity. Official MiniMax release post. Accessed: 2026-04-18. find
  12. OpenAI, :, Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K. Arora, Yu Bai, Bowen Baker, Haiming Bao, Boaz Barak, Ally Bennett, Tyler Bertao, Nivedita Brett, Eugene Brevdo, Greg Brockman, Sebastien Bubeck, and 108 others. 2025. gpt-oss-120b & gpt-oss-20b model…
  13. OpenAI. 2025. Introducing GPT-5 for developers. Official OpenAI release. Accessed: 2026-04-18. find
  14. OpenRouter. 2025. Openrouter: Unified api for llms and automatic routing. https://openrouter.ai. Accessed 2026-01-03. ↗
  15. Long Phan and et.al. Alice Gatti. 2025. Humanity’s last exam. Preprint, arXiv:2501.14249.
  16. Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, and 1 others. 2023. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789.
  17. Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36:68539–68551.
  18. Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, and 1 others. 2025a. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534.
  19. Tongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, and 1 others. 2025b. Tongyi deepresearch technical report. arXiv preprint arXiv:2510.24701. — [arXiv:2510.24701](https://arxiv.org/abs/2510.24701) find
  20. Qianxin Wang, Liqi Chen, Souradip Chakraborty, Qing Dong, Haoran Zhang, Yang Liu, Zhenghai Qi, Tianyi Guo, and Meng Liu. 2024. Embedllm: Learning compact representations of large language models. arXiv preprint arXiv:2401.11623. — [arXiv:2401.11623](https://arxiv.org/abs/2401.11623) find
  21. Ruoyao Wang, Peter Jansen, Marc-Alexandre Côté, and Prithviraj Ammanabrolu. 2022. Scienceworld: Is your agent smarter than a 5th grader? arXiv preprint arXiv:2203.07540. find
  22. Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh J Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, and 1 others. 2024. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. Advances in Neural Information Processing Systems, 37…
  23. John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024. Swe-agent: Agent-computer interfaces enable automated software engineering. Advances in Neural Information Processing Systems, 37:50528– 50652.
  24. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. In The eleventh international conference on learning representations.
  25. Haozhen Zhang, Tao Feng, and Jiaxuan You. 2025a. Router-r1: Teaching llms multi-round routing and aggregation via reinforcement learning. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. find
  26. Yiqun Zhang, Hao Li, Jianhao Chen, Hangfan Zhang, Peng Ye, Lei Bai, and Shuyue Hu. 2025b. Beyond gpt-5: Making llms cheaper and better via performance-efficiency optimized routing. In Proceedings of the 2025 7th International Conference on Distributed Artificial Intelligence, pages 122–129. find
  27. Yiqun Zhang, Hao Li, Chenxu Wang, Linyao Chen, Qiaosheng Zhang, Peng Ye, Shi Feng, Daling Wang, Zhen Wang, Xinrun Wang, and 1 others. 2025c. The avengers: A simple recipe for uniting smaller language models to challenge proprietary giants. arXiv preprint arXiv:2505.19797. — [arXiv:2505.19797](https://arxiv.org/abs/2505.19797) find
  28. Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, and 1 others. 2023. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-mtrouter-cost-aware-multi-turn/