註冊並分享邀請連結,可獲得影片播放與邀請獎勵。

DAIR.AI
@dair_ai
Democratizing AI research, education, and technologies. Learn about AI Agents for FREE at
加入 July 2017
1 正在關注    134.4K 粉絲
Banger paper from Meta Superintelligence Labs on user simulators for agent training. Agent RL setups usually let an assistant LLM play the user, so the simulated user is too cooperative and too explicit. A fixed GPT-5.5 agent finds tau-bench tasks easier with these users than with real people. This work trains MIMESIS, a 9B user simulator, on human conversations and 13 behavior patterns observed in real users. It beats Claude Opus 5 on behavioral fidelity by 13.4 points. Agents trained against it outperform agents trained against GPT-5.5 under all nine user simulators they never saw. They find that adding a coaching step that turns the simulator's private reasoning into feedback adds further gains. Paper:
顯示更多
0
24
205
23
轉發到社區