Canonical paper record
EAU
arXiv:2610.16523
EAU: Environment-Agent-User Co-evolution for Training User-Centric Agents is a meta-agent framework for user-centric agents.
It adapts a hybrid user simulator, environment feedback, reward design, and RL configuration within and across runs. Evaluation covers UserGym, ColBench, and τ²-bench with Qwen3-4B and Qwen2.5-7B-Instruct.
Official project site: eau-meta-agent.github.io.
01 / Authors and affiliations
A six-author
collaboration.
Min Cai — University of Alberta; Dan Zhang — National University of Singapore; Gao Xing and Libin Ma — Kuaishou Technology; Zhaohui Xu — Independent Researcher; Xi Ye — University of Alberta and Amii.
02 / Reported evidence
Benchmarks and
results.
On Qwen3-4B, EAU reports 0.623 on UserGym, 0.600 ColBench task success, and 0.627 overall. The ColBench task-success result is a 36.7% relative gain over InfoPO under the paper’s protocol.