We propose Rectified Policy Optimization (RePO) to mitigate "safety compensation", which replaces the average safety metric with stricter safety constraints.

Creating safe and reward maximization policies from offline data via min-max optimization formulation and solving it using no-regret algorithms

We view offline optimization from the new lens of a distributional translation task which can be modeled with a generalized Brownian Bridge Diffusion process mapping between the low-value and high-value input distributions.

Learning Reconfigurable Representations for Multimodal Federated Learning with Heterogenous Missing Data