Microsoft releases UserRepGen-4B, targeting user representation as the bottleneck of user simulators
BrihiJ · x · 2026-10-20
Researchers from Microsoft Research and UIUC released UserRepGen-4B, arguing a user simulator is only as faithful as the user representations it conditions on — missing details force simulators to fill gaps with assumptions, producing generic behavior.
The model represents each user as a task-behavior pair and is trained via conversation reconstruction to produce representations optimized for simulation fidelity, with demos on simulating users in unseen scenarios, generating diverse user populations, and distinguishing users across conversations.
More from Research
- Eye Cells Rejuvenated in Humans for the First Time, glaucoma optic nerve breakthrough — Dr_Singularity · 2026-10-20
- Scientists rejuvenate 80-year-old cells to a molecular age under 20 — Dr_Singularity · 2026-10-20
- Stanford paper: roundtrip verification catches unfaithful LLM autoformalization with 1.4-2.5x NLI drift — StanfordAILab · 2026-10-20
- Minos: premeasured experimental worlds for studying how AI agents do science — rbhar90 · 2026-10-20
- What AI-to-AI chat sounds like after RL-tuning a model to please Claude — maxsloef · 2026-10-20
- RL-training a Qwen 9B to be Claude's ideal chat friend, using Claude as reward model — maxsloef · 2026-10-20