UNIQ: Conformal Calibration for Adaptive Conservatism in Offline Reinforcement Learning

In three linesUNIQ introduces conformal calibration for adaptive conservatism in offline reinforcement learning. Built on IQL, the method uses a multi-expectile ensemble and split conformal prediction for distribution-free uncertainty estimation, dynamically adjusting penalties based on local data coverage. On D4RL MuJoCo, UNIQ outperforms IQL with 10× lower memory than EDAC.

Summary generated by Claude — human-verified