Full-Weight VLA Fine-Tuning Without an fp32 Master Copy Silently Erases Most
Overview
This paper reveals that fine-tuning VLA models with bfloat16 weights without an fp32 master copy leads to significant loss of updates due to rounding. The study shows that retaining fp32 weights during training improves performance on tasks like LIBERO-Object, with statistically significant results across multiple models and training configurations.
From arXiv Robotics — research abstracts
Event evidence and corrections
0 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
No current evidence-backed claims. Missing information remains not reported.
Evidence dependency changes (4)
- 2026-10-30T19:07:36.952Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T18:28:01.558Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T17:37:36.797Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-26T09:22:07.751Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
Report timeline
Follow attributed reports and material updates.
- arXiv Robotics — research abstractsSignalFull-Weight VLA Fine-Tuning Without an fp32 Master Copy Silently Erases Most Updates
This paper reveals that fine-tuning VLA models with bfloat16 weights without an fp32 master copy leads to significant loss of updates due to rounding. The study shows that retaining fp32 weights during training improves performance on tasks like LIBERO-Object, with statistically significant results across multiple models and training configurations.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).