Full-precision weight copies help two robot models learn tasks
Full-Weight VLA Fine-Tuning Without an fp32 Master Copy Silently Erases Most Updates · arXiv:2610.31859
In plain words
When a vision-and-language model learns a new robot task, its training makes many small changes to its internal weights. This paper studies what happens when those weights are stored in a 16-bit format: some changes are too small to survive rounding, so the stored weights do not change.
The author tested keeping a full-precision copy of the trainable weights while still computing in 16-bit. On one simulated task suite, two tested models scored higher with the copy in the reported comparisons. The tests covered only those models and that suite.
The research
When a vision-and-language model is trained further to control a robot, can small training updates disappear if its trainable weights are stored in a lower-precision format?
If many updates leave the stored weights unchanged, does keeping a more precise copy improve the model’s measured success on a new task?
The paper measures erased weight updates and compares keeping trainable weights in full precision with bf16 storage in two VLA models.
The author probes SmolVLA’s stored weights during training, then compares task success with and without a full-precision copy. The probe finds most counted elements unchanged after sampled steps in the tested bf16-storage runs.
On LIBERO-Object, the authors’ scorer reports higher mean success with the copy for SmolVLA and pi0.5 across the tested learning rates. A SmolVLA control against all-bf16 storage also favors the copy at the lower rate.
Together, the weight probe and task comparisons support the conclusion that bf16 storage without a copy can hinder these tested fine-tuning runs.
The task-success tests cover only SmolVLA and pi0.5 on one LIBERO suite, with few seeds and shortened training schedules.
They do not separate erased weight updates from the effects of storing AdamW’s moments in bf16, so the size and cause of any gain in other models, tasks, or schedules remain unknown.
What the evidence shows
- When training updates vanish
The paper examines a specific risk during vision-language-action (VLA) fine-tuning: a model’s trainable weights may be stored in bfloat16 (bf16), a 16-bit format, while an optimizer writes updates directly to them. The central question is whether rounding can erase enough small changes to matter when the model is trained further on a new task. The author also checks whether common open VLA training recipes use this storage arrangement.
- From rounding to stalled weights
The paper’s analysis shows why direct storage matters: every update smaller than 1/512 of a weight’s magnitude rounds back to that weight, and some larger updates can also disappear. To see whether this happens during actual runs, a probe compares stored trainable weights before and after sampled optimizer steps.
In SmolVLA runs without a full-precision copy, it found 79–99% of counted weight elements unchanged per sampled step at a learning rate of 10^-4, and 96–99.95% at 10^-5.
- Testing task performance
The author then tested whether keeping trainable weights in 32-bit floating point changes task success, while computation still used bf16. In the main SmolVLA comparison, the settings shared a starting checkpoint, demonstrations, optimizer, schedule, learning rate, and evaluation conditions; their trainable-weight storage differed.
On LIBERO-Object, the authors’ scorer measured success across attempts and the reported figures average three training seeds. The copy had higher mean success than LeRobot’s recipe at both tested learning rates.
- The pattern across two models
A second model, pi0.5, showed the same direction of comparison in the tested settings. With a shortened 2,500-step schedule from the general base model, the copy’s mean success was 97.7% versus 89.0% for LeRobot’s bf16-flag recipe at the default learning rate, and 80.9% versus 5.1% at a ten-times lower rate.
These are two-seed means on the same simulated task suite, not a guarantee for other training setups.
- Controls and boundaries
The SmolVLA all-bf16 control also scored below the copy at the lower learning rate, so the finding is not confined to comparison with LeRobot’s recipe. Yet the gap depended on how the trained model was evaluated: replanning every 10 actions reduced the SmolVLA difference to 4.2 points, compared with 10.2 points using open-loop chunks of 50 actions.
The tests cover two models, one suite, few seeds, and shortened schedules, and do not isolate erased updates from bf16 optimizer moments.
Deep
- Original paper authors
- Shubham Singh
- Submitted to arXiv
- License
- arXiv non-exclusive