Most VLAs lose their VLM's vision-language skills during pretraining

DJiafei · x · 2026-11-16

Research by Ville Kuosmanen shows most vision-language-action (VLA) models lose the vision-language skills of their underlying VLM during pretraining. By isolating the VLM weights of open-source VLAs and evaluating them on vision-language benchmarks, the authors find three distinct categories of models emerge, highlighting how VLA training erodes general multimodal understanding.

Related event: Most VLAs Lose Their Vision-Language Skills After Pretraining(2 posts)→

Original post →

More from Embodied

Embodied channel →