A human-trained user simulator improves agent training
In plain words
A simulated user helps researchers train an agent for conversations that unfold over several turns. But a language model pretending to be a user may answer too clearly or cooperate too readily, changing how difficult the task seems.
The paper introduces MIMESIS, trained on human conversations and recurring behavior patterns. Agents trained with it scored higher than agents trained with GPT-5.5 when tested with nine other simulated users. The study does not test whether those agents perform better with real people.
The research
When agents learn from simulated conversations, how can we tell whether those conversations preserve the difficulty people create? A fixed GPT-5.5 agent succeeded in 63.6% of interactions with real users, but 82.4–84.4% with frontier API simulators. The contrast shows that the simulated user can change measured task success.
MIMESIS learns user behavior from human conversations, then supplies simulated interactions and retrospective coaching for agent training.
For a Qwen3-8B agent, replacing GPT-5.5 with MIMESIS-9B as the simulated training user raised mean task score from 26.10 to 29.54 across eight environments and nine evaluation simulators.
Adding Coached On-Policy Self-Distillation (CSD), which uses simulator feedback to coach training, raised the mean to 31.09. Scores rose under all nine users in both comparisons, supporting transfer among simulated users.
The agent results measure transfer across nine language-model simulators, not performance with human users. The coaching notes also come from the simulator’s generated thoughts and reactions; they are not observations of real users’ mental states.
Analysis
Deep
- Authors
- Hoang Phan, Dat Huynh, Andrey Zhmoginov, Qi Zeng, Wancen Mu
- License
- CC0