EAU co-evolves environment, agent, and user during RL
Original title: EAU: Environment-Agent-User Co-evolution for Training User-Centric Agents
Authors report a 14.7–15.0% score gain over the strongest baseline and up to 36.7% more ColBench completions.
EAU is a meta-agent framework that changes the user simulator, environment, reward design, and RL configuration while an LLM agent is trained. It monitors trajectories, tool traces, rewards, and training statistics, diagnoses failures, and hot-swaps validated edits into the running job. User and environment dynamics use an editable symbolic representation, while the agent remains a neural LLM trained with RL. Across three interactive-agent benchmarks and two unnamed backbone models, the authors say EAU beats prior RL methods.
Unconfirmed: Gains are author-reported; backbone models and the other two benchmarks are not named.
- Adaptations persist across runs instead of restarting from a fixed setup.
- The paper says simulated users and tools can be inconsistent or unreliable.
- ColBench successful task completion rises by up to 36.7%, per the authors.
- Overall score improves 14.7–15.0% over the strongest baseline.
From the original article
1 min read at huggingface.co
Training LLM agents to interact with users over multiple turns is challenging because the learning signal depends not only on the agent, but also on the user simulator and environment, which jointly shape the agent's experience. Simulated users can be stochastic or inconsistent, while environments may contain unreliable tools, incomplete state coverage, and incorrect verifiers.
By Min Cai, Dan Zhang, Gao Xing · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.