Fan Feng

GPT-Policy is introduced, a general-agent framework for in-context robot learning that integrates a context compiler that preserves task-relevant visual transitions, a VLM that proposes robot-tool actions, and a constrained controller that verifies and executes each action and reports its outcome.

Interaction-weighted Resampling transforms how robots learn to manipulate objects, yielding a 19.8% boost in performance and a breakthrough in real-world goal achievement.

Achieving high-fidelity cloud removal, IB-HFN outperforms traditional methods by effectively decoupling noise suppression from texture preservation.

Foundation model embeddings contain so much irrelevant detail that projecting them into a task-centric latent space actually *improves* world model quality and control performance.

Learn to detect unknown network attacks by explicitly modeling what they are *not*.

World models can now self-improve by identifying their own prediction errors, thanks to a clever decomposition of action-conditioned prediction into easier-to-verify components.

Stop reward hacking: disentangling causal and non-causal factors in reward models makes RLHF more robust.