Featured

Study: Full-weight VLA fine-tuning without an fp32 master copy silently erases most updates

The paper points out that when VLA models are fully fine-tuned in bf16 without keeping an fp32 master copy, AdamW rounds away any update smaller than 1/512 of a weight; 9 of 12 open-source VLA codebases train this way by default or in documented recipes, including LeRobot's SmolVLA default configuration and the π0/π0.5 guides.

Part of the event: Study: bf16 fine-tuning of VLA silently drops updates

This page provides a summary and source link only.