The authors first finetune an LLM on user turns from a corpus of user-assistant interactions, then post-train it with multi-turn reinforcement learning.
Reinforcement learning yields user simulators that stay faithful and sound natural
The authors' UserLM-2.0 model produces valid simulated conversations 46.2% of the time, over three times the rate of the strongest prior simulator.
Big Tech
Mustafa Omer Gul · Philippe Laban · Tanya Goyal · Jennifer Neville
Cornell University · Microsoft Research
Research Digest··3 min read
The authors address a trade-off in user simulators for evaluating AI assistants: prompted LLMs follow complex intents but sound unnatural, while specialized simulators sound human but drift from task instructions.
§