Reinforcement Learning

Large Language Models

StaRPO: Stability-Augmented Reinforcement Policy Optimization

🎉 Accepted at NeurIPS 2026! Rewarding LLMs for reasoning that stays coherent and goal-directed, not just for landing the right final answer, makes their reasoning both more …

Jinghan Zhang
•
Read more
User Simulation

HumanLM: Simulating Users with State Alignment Beats Response Imitation

🎉 Accepted at ICML 2026! Training user simulators to infer what real people believe and feel, not just mimic how they write, produces far more realistic simulated users.

Shirley Wu
•
Read more