Accepted to ACL 2026 Main Conference (Oral)
With the rise of the Agent Web and Model Context Protocol (MCP), the agent ecosystem is evolving into an open collaborative network, exponentially increasing accessible tools. However, current architectures face severe scalability and generality bottlenecks.
We propose ACE-Router, a pipeline for training history-aware routers to empower precise navigation in large-scale ecosystems. By leveraging a dependency-rich candidate graph to synthesize multi-turn trajectories, we effectively train routers with dynamic context understanding to create the plug-and-play Light Routing Agent.
- Self-Evolutionary Graph Construction -- Expands and structures the candidate space via mutation and relation modeling.
- Multi-Agent Simulation -- Synthesizes interaction trajectories to extract history-aware supervision signals.
- Light Routing Agent -- A plug-and-play module that seamlessly integrates the trained router into existing agent pipelines.
- Cross-domain Transferability -- A router trained solely on tool data generalizes to multi-agent collaboration with minimal adaptation.
- Robustness & Scalability -- Maintains exceptional robustness against noise and scales effectively to massive candidate spaces.
- [2026-05-13] We released the code and paper for ACE-Router.
- [Coming Soon] Data and model weights will be released. Stay tuned!
The inference and evaluation code is now available under AceRouter/. See AceRouter/README.md for setup, the smoke test, and how to plug in your own router model.
cd AceRouter
conda create -n toolace python=3.11 -y && conda activate toolace
conda install -c conda-forge nodejs -y
pip install -e .- MCP-Universe: For the evaluation framework and benchmarks.
- MCPMark: For the evaluation benchmarks.
- LLaMA-Factory: For the SFT implementation.
- Qwen: For the powerful base models.
@misc{acerouter2026,
title={ACE-Router: Generalizing History-Aware Routing from MCP Tools to the Agent Web},
author={Zhiyuan Yao and Zishan Xu and Yifu Guo and Zhiguang Han and Cheng Yang and
Shuo Zhang and Weinan Zhang and Xingshan Zeng and Weiwen Liu},
year={2026},
eprint={2601.08276},
archivePrefix={arXiv},
}