Distilling Instructions into Weights: Closed-Form Edits Fix a Robot Policy's Misreading
Distilling Instructions into Weights: Closed-Form Edits Fix a Robot Policy's Misreading
This paper introduces a method to repair instruction-following failures in generalist robot policies by directly editing model weights using closed-form distillation of explicit rewordings, avoiding gradients or new training data. The authors demonstrate that this approach significantly improves task success rates on held-out seeds compared to random edits or no edit, while releasing an open-source tool for end-to-end reproduction of the fix.
The authors propose a closed-form weight-editing method to correct specific instruction-following failures in a zero-shot vision-language-action policy without retraining or gradient updates, reporting success rate improvements on held-out seeds and releasing an end-to-end tool for reproduction.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
success episodes
27–106 episodes
Reported trials
View original evidence
e before it reaches the bowl. Replaying 21 of these decisions under rewordings shows that it binds "in the bowl" to the wrong object. Closed-form instruction distillation copies the per-layer effect of an explicit rewording into the language model's MLP weights. On 128 held-out seeds, success rises from 27 to 106 episodes, with 81 fixed and 2 broken ($p=7\times10^{-22}$), while an update of the same size in random di
Open source S3success episodes
8
Reported trials
View original evidence
rections leaves success unchanged. The edit also lifts two other tasks without detectable harm on the remaining 13, and a variant that stores 0.16% of the parameter count keeps most of its effect. Placed in the action expert or action head instead, the same distillation helps less or not at all. Our released tool reproduced the fix end to end on new held-out seeds, from 8/64 to 36/64. Edits aimed at perception, a mis
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
e before it reaches the bowl. Replaying 21 of these decisions under rewordings shows that it binds "in the bowl" to the wrong object. Closed-form instruction distillation copies the per-layer effect of an explicit rewording into the language model's MLP weights. On 128 held-out seeds, success rises from 27 to 106 episodes, with 81 fixed and 2 broken ($p=7\times10^{-22}$), while an update of the same size in random di
Open source S3
rections leaves success unchanged. The edit also lifts two other tasks without detectable harm on the remaining 13, and a variant that stores 0.16% of the parameter count keeps most of its effect. Placed in the action expert or action head instead, the same distillation helps less or not at all. Our released tool reproduced the fix end to end on new held-out seeds, from 8/64 to 36/64. Edits aimed at perception, a mis
Open source S4
sing grasp, contact physics, or the gripper threshold did not help, and of seven blind Claude Opus editors, only the one given the misread instruction produced an edit that passed held-out scoring. These results suggest that weight edits work best when a known-good input already exists to distil. Because reruns with unchanged weights flip about a third of paired outcomes, evaluations should pair seeds, rerun the uned
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org