Opening the policy, not the prompt

Interpretability spent years explaining. It is now being used to edit — and robotics is where that transition pays first, because data is expensive enough to make surgery worth attempting.

Work on mechanistic finetuning of vision-language-action models proposes using interpretability findings to make targeted edits from few-shot demonstrations, rather than retraining on more examples.

Why robotics first

Because the economics are different. When collecting a thousand more demonstrations costs weeks of robot time, a technique needing five examples is worth enormously more than the same method applied to text, where you can simply gather more data.

It is also interpretability finally being used as an instrument rather than an explanation, which is the transition the field has been promising for years. Locate the mechanism, change it directly, instead of hoping gradient descent finds it.

The control-theory pairing

Combining mechanistic interpretability with optimal control is a more natural fit than it first appears. Control theory is the mature discipline for steering a system toward a trajectory given a model of its dynamics. Interpretability is what supplies the model of the dynamics.

The appeal is that it targets robustness rather than accuracy. A policy that performs well on the test distribution and degrades unpredictably outside it is the standard failure, and average-case training does nothing about it.

The caveat that never goes away

Steering demonstrates a direction in activation space is causally relevant. It does not establish that it is the only such direction, or that suppressing it removes a capability rather than one expression of it.

Which is the same limit that shows up everywhere in this literature: a mechanism located on the cases you tested tells you about those cases. Coverage against inputs nobody tried is, by construction, unmeasured — and coverage numbers, not example circuits, are what a safety claim actually needs.

arXiv — Mechanistic finetuning of vision-language-action models via few-shot demonstrations → · arXiv — Understanding emergent misalignment via feature superposition geometry → · Anthropic — Tracing the thoughts of a large language model →