Most VLAs lose their base VLM's vision-language skills during robot pre-training, eval shows

DJiafei · x · 2026-11-16

By isolating the VLM weights inside open-source VLAs and benchmarking them on 8 vision-language tasks, an analysis shows most VLAs dramatically degrade their base model's skills after robot pre-training: OpenVA drops 50.3 pts vs Prismatic 7B, NVIDIA GR00T N1.7 drops 65.3 pts vs Cosmos-Reason2, and G0.5 drops 33.5 pts vs Qwen3.5-2B-Base. Exceptions exist: π0-5 gains +4.1 pts over PaliGemma 3B, and MolmoAct2 loses only 1.5 pts. Models fall into three categories: near-total loss, partial retention, and near-full retention—highlighting a systemic capability-erosion problem in the VLA pipeline.

Original post →

More from Embodied

Embodied channel →