Yihao Liu

Treating all latent tokens identically bottlenecks multimodal test-time compute—disentangling hidden-state updates into visual-grounding and high-entropy reasoning channels beats standard output-space search under matched sample budgets.

Chasing large token-level teacher-student gaps in agent distillation is largely misguided: local divergence fails to predict task completion, whereas unpenalized minor deviations often trigger catastrophic multi-turn failure.

Current video generation models fall short of capturing the full distribution of possible behaviors, revealing a critical gap in probabilistic alignment that needs to be addressed.

WorldRover turns long-horizon world exploration into a scalable data-generation problem, providing a treasure trove of richly annotated sequences for training AI models.