Most VLAs Lose Their Vision-Language Skills After Pretraining

A study by Ville Kuosmanen et al. finds that most vision-language-action models lose much of their underlying VLM's vision-language skills after robot pretraining, as shown by evaluating isolated VLM weights on eight vision-language benchmarks.

2026-11-16 ~ 2026-11-16 · 2 related posts