← Azimuth Daily
Capturing data…
Capturing data…
TUE, 27 OCT · 97 ITEMS
"Full-weight VLA fine-tuning without an fp32 master copy silently erases most updates" — arXiv cs.RO · AI & Technology
This paper reports that fine-tuning a Vision-Language-Action () model on all its weights (full-weight) without keeping a separate 32-bit float () copy of the weights silently destroys most of the gradient updates. In practice, using only precision for both forward and backward passes causes updates to be lost due to limited precision; experiments on the robotics benchmark show at least 79% of updates are erased, and 99% of models on the Hugging Face Hub that were fine-tuned this way show effectively no learning.
The claim comes directly from the primary source—the arXiv paper itself (arXiv:2610.31859). The search results confirm the paper's existence and the specific quantitative findings (79% erasure in LIBERO, 99% of Hub models affected). As a self-contained research result, the claim is solid, though it represents a single study that would benefit from replication.
The paper was published on arXiv on 2026-10-26, so it is genuinely new research.