Mian Wu

We propose a post-training methodology for improving language models on open-ended generation tasks. Instead of relying on static reward models, RLAC trains a dynamic LLM critic alongside the generator using adversarial gameplay. The critic identifies the most likely failure modes which are then validated externally, adapting as the generator improves. We demonstrate improvements in factual accuracy for text generation and correctness for code generation across multiple benchmarks.

We introduce a demo-free reward framework that leverages pretrained video diffusion models to generate goal-conditioned reward signals for RL agents, eliminating the need for hand-crafted reward design or expert demonstrations.