bf16 breaking VLA fine-tuning sparks debate: fp32 or stochastic rounding already known?

giffmana · x · 2026-11-18

mwulfmeier calls the paper on bf16 breaking VLA fine-tuning one of the month's most interesting; giffmana pushes back that keeping main weights in fp32 or using stochastic rounding has been well established across domains for a while. The exchange highlights low-precision pitfalls when fine-tuning robot policies and known mitigations.

Related event: Paper Claims bf16 Breaks VLA Fine-tuning, Sparking Debate(2 posts)→

Original post →

More from Embodied

Embodied channel →