Mechanistic finetuning: editing a robot policy with interpretability rather than data
Work on mechanistic finetuning of vision-language-action models from few-shot demonstrations proposes using interpretability findings to make targeted edits, rather than retraining on more examples.
This is interpretability being used as an instrument rather than as an explanation, which is the transition the field has been promising for several years. If you can locate the mechanism responsible for a behaviour, you can in principle change it directly instead of hoping gradient descent finds it.
Robotics is a sensible place to attempt it because data is the binding constraint. When collecting a thousand more demonstrations costs weeks of robot time, an edit that needs five examples is worth a great deal more than the same technique applied to text.
The standing caveat applies. Locating a mechanism for a behaviour on the cases you tested does not establish that it is the only mechanism, or that the edit generalises to inputs nobody tried.
arXiv — Mechanistic finetuning of vision-language-action models via few-shot demonstrations → · arXiv — Mechanistic interpretability for neural networks: circuits, sparse features and symbolic reasoning → · Anthropic — Tracing the thoughts of a large language model →