Within-run adaptation
EAU can revise supported user rules, environment feedback, rewards, and RL configuration while a training run continues.
Research context
Where does EAU
fit?
EAU studies agent-environment-user co-evolution for LLM agents: an adaptive training system jointly revises the environment, the agent training process, and a simulated user.
The method targets multi-turn reinforcement learning with users and tools. It treats the training system as editable, using rollout evidence to select and validate interventions while evaluation remains fixed.
01 / Mechanism
Co-evolution at
two timescales.
Across-run adaptation
It consolidates evidence from earlier runs, restarts when necessary, and uses accumulated experience to guide later interventions.
Three coupled components
The agent policy learns while the environment and hybrid user simulator are adapted from observed training failures.
02 / Evidence boundary
Training evolves.
Evaluation stays fixed.
EAU is evaluated on UserGym, ColBench, and τ²-bench with Qwen3-4B and Qwen2.5-7B-Instruct. Final-test feedback stays outside the adaptation loop. Reported gains describe these benchmark settings and do not establish performance for every agent environment.