Reproduce “EvoMCP: Co-evolving MCP agents and training data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xuanang Gao, Kepeng Lin, Xinyan Shi, Yong Deng, Shuheng Shen, Changhua Meng, Xiaofeng Wu, Jie Yang and 1 more
EvoMCP adapts tool-using assistants’ training data to student behavior by evolving verified executable tasks and repairing or refining teacher demonstrations. Across three rounds, it trains on accepted task–trajectory pairs, using benchmark feedback to select learning targets. On Toolathlon, Qwen3.6-27B reaches 53.4% task success versus 39.8% for matched-step Static SFT; this is benchmark-guided, not held-out generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.