Lang Feng

SEED transforms past experiences into actionable skills, allowing reinforcement learning policies to evolve and improve in real-time.

A single representation can be easily manipulated, but a balanced ensemble of encoders dramatically enhances image generation quality and robustness.

REAR transforms how we achieve user preference alignment in LLMs, enabling scalable realignment without costly retraining.

OPID achieves a remarkable boost in agent performance by leveraging hierarchical skills extracted from on-policy trajectories, transforming sparse rewards into dense, actionable insights.

Don't let valuable steps in failed trajectories go unnoticed: GraphGPO leverages state-transition graphs for fine-grained credit assignment in agentic RL, boosting performance and efficiency.

Generate minute-long, high-fidelity animations without visual degradation or character drift using a surprisingly simple latent flow restoration technique.

Visuomotor control can now generalize to unseen environments and instructions by grounding world models in a vision-language latent space, outperforming standard vision-language approaches by a large margin.

Context inconsistency in stepwise group-based RL can severely bias advantage estimation, but a hierarchical grouping strategy can fix it without extra compute.