Jong Chul Ye

Discarded tail actions in chunked robot policies remain faithful to counterfactual replans until action velocity fluctuates, allowing policies to safely extend execution horizons and slash inference calls by up to 1.7× at zero extra cost.

Online critic backpropagation for flow-based robot policies is completely unnecessary: amortizing trajectory-aware adjoint costates into a lightweight network achieves superior value alignment at 3.6× the speed.

By injecting alignment guidance directly into the diffusion model's score function, AGSM avoids the pitfalls of contrastive learning and generates images with improved counting accuracy and semantic fidelity.

Generate videos several times longer than the native window length without training, simply by intelligently stitching together overlapping windows with manifold constraints.

Controllable 3D generation takes a leap forward with 3D-ReGen, a framework that leverages an initial 3D shape for tasks like enhancement and editing, outperforming existing methods.

LLMs can help diagnose rare diseases from X-rays, even ones the model has never seen before.