Jun Li

Predicting compact, learned future representations for world-action modeling slashes training costs by 65x and outperforms native visual targets, even with smaller models.

Semantic relevance is a dangerously loose proxy for safety: replacing standard RAG with formal, scene-grounded rule entailment prevents vision-language planners from acting on ungrounded risk assumptions.

JaxPIP is introduced, a JAX-based implementation that reformulates PIP/FI evaluation into tensorized linear algebra operations that allows efficient execution with just-in-time compilation and enables large-scale batch evaluation of energies and forces.