Taesup Kim

By factoring transition extent and mode in latent actions, PoLAR achieves superior policy performance, revealing the critical role of latent action geometry in robot learning.

VLA models can suffer drastic performance drops due to joint-level faults, even when actions are physically feasible.

Hallucinations aren't always due to a lack of knowledge; larger LLMs often "know" the right answer but fail to commit to it, scattering probability across multiple surface forms.