Full-Weight VLA Fine-Tuning Without an fp32 Master Copy Silently Erases Most Updates
arXiv:2610.31859v1 Announce Type: new Abstract: When a vision-language-action (VLA) model is fine-tuned with its trainable weights in bfloat16 (bf16) and no fp32 master copy, AdamW writes each update straight into bf16, silently rounding away any update smaller than 1/512 of a weight. In 9 of 12…
Read original reporting ↗This Observatory page contains only an indexed headline and short summary. It does not reproduce the article or substitute for the publisher’s reporting, corrections, context or terms.
Read the original at arXiv cs.RO ↗