Hybrid user simulator
Combines symbolic rules with natural LLM responses. EAU can revise disclosure rules, user personas, generation settings, and simulator mixing.
Method summary
What does EAU
adapt?
EAU is a meta-agent framework that jointly adapts a hybrid simulated user, environment feedback, reward design, and reinforcement-learning configuration while training user-centric agents.
It uses trajectories, tool traces, and training statistics to diagnose the interaction system and validate changes before applying them.
01 / Editable components
The training system becomes
part of the optimization.
Environment feedback
EAU can revise environment behavior, tool interaction feedback, and the information returned to the policy.
Reward design and RL configuration
Rewards and reinforcement-learning configuration can change as the policy and its failure modes change.
02 / Timescales
Within a run.
Across runs.
Within a run, supported interventions can be applied while training continues. Across runs, validated experience is consolidated and reused in later training attempts.
03 / Evaluation
Interactive benchmarks and
two policy backbones.
EAU is evaluated on UserGym, ColBench, and τ²-bench with Qwen3-4B and Qwen2.5-7B-Instruct. On Qwen3-4B, it reports 0.600 ColBench task success, a 36.7% relative gain over InfoPO.