Elamur Research⌕
☰
home / papers / 2026 / ACE-Router: Generalizing History-Aware Routing from MCP Tools to the Agent Web

ACE-Router: Generalizing History-Aware Routing from MCP Tools to the Agent Web

Paper · 2026 · Zhiyuan Yao, Zishan Xu, Yifu Guo et al.

Summary

As the Model Context Protocol (MCP) and the emerging Agent Web expose ever-larger pools of tools and agents, picking the right one becomes a bottleneck. Listing every schema in the prompt runs out of context window. Embedding-based retrieval cannot separate near-duplicate tools and ignores what happened earlier in the conversation.

ACE-Router addresses this with a pipeline for training routers that condition on both the query and the dialogue history. It works in three stages:

  • Graph-based expansion. A candidate graph links tools by similarity or dependency, and self-evolutionary mutation adds functionally close variants.
  • Trajectory synthesis. Random walks over this graph choose tool subsets. A multi-agent simulation then turns these subsets into multi-turn trajectories that serve as routing supervision.
  • Deployment. The trained router is packaged as the plug-and-play Light Routing Agent.

On MCP-Universe and MCP-Mark, ACE-Router beats embedding retrievers, ReAct agents and strong LLM-based routers. It stays stable as noisy or real external tools are added, and it transfers to agent selection without extra fine-tuning. This makes it a practical basis for large-scale orchestration.

Key points

  • ACE-Router reaches 53.44% accuracy on MCP-Universe and 60.00% on MCP-Mark in the clean setting. The best LLM router baseline on MCP-Universe is Gemini-2.5-Pro at 49.79%.
  • Under the Multi, +Mutation and +LiveMCP settings, ACE-Router's MCP-Universe accuracy stays roughly between 52.6% and 53%. Embedding retrievers fall to 20–32% once LiveMCP tools are added.
  • Training without dialogue history lowers accuracy from 53% to 48% on MCP-Universe and from 60% to 52% on MCP-Mark. The authors attribute this to lost sequential-dependency reasoning and lost error recovery.
  • Mutation enlarges the tool pool from 627 to 2,005 candidates. Removing mutation gives 47.12% / 44.00%, and vanilla fine-tuning gives 46.70% / 42.00% (MCP-Universe / MCP-Mark).
  • The Light Routing Agent uses only two interface tools, Router Invocation and Execution. Code is available at github.com/euyis1019/ACE-Router.

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 Jiao Tong UniversityShanghai · China212 works · 19,544 citations🎓Nanyang Technological UniversitySingapore · Singapore184 works · 27,484 citations🎓Zhejiang UniversityHangzhou · China224 works · 13,587 citations🏢Huawei TechnologiesShenzhen · China96 works · 15,432 citations🎓Sun Yat-sen UniversityGuangzhou · China72 works · 9,502 citations🎓Hangzhou Dianzi UniversityHangzhou · China15 works · 1,061 citations

 Fields, topics and concepts

📚 Shelves 5The 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 worksInformation Systems · 1,730 works

Cite

Cite as Zhiyuan Yao, Zishan Xu, Yifu Guo, Zhiguang Han, Cheng Yang, Shuo Zhang, Weinan Zhang, Xingshan Zeng and Weiwen Liu. 2026. ACE-Router: Generalizing History-Aware Routing from MCP Tools to the Agent Web. 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.281

@inproceedings{yao2026ace,
  title = {{ACE-Router: Generalizing History-Aware Routing from MCP Tools to the Agent Web}},
  author = {Zhiyuan Yao, Zishan Xu, Yifu Guo, Zhiguang Han, Cheng Yang, Shuo Zhang, Weinan Zhang, Xingshan Zeng and Weiwen Liu},
  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.281},
  url = {https://doi.org/10.18653/v1/2026.acl-long.281}
}

References

2 references, 1 of them with a card of ours. ↗ leads to the work.

  1. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.
  2. Anthropic. 2024. Introducing the Model Context Protocol. https://www.anthropic.com/news/ model-context-protocol. \#\# Conclusion ↗

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yao-2026-ace-router-generalizing-history-aware/