Certified Mechanistic Edits: Behavioral Guarantees for Skill Removal and Preservation
Researchers certify that mechanistic edits provably remove one neural network skill while preserving another, replacing test-only validation with formal behavioral guarantees.
The paper certifies the behavioral effect of mechanistic edits used for unlearning (ablations, weight edits, activation steering): disabling a circuit removes one skill and provably preserves another for every input in a continuous region, a feature non-interference guarantee. Sound bound propagation reaches roughly 9x the input-perturbation dimension of exact solvers, demonstrated from toy ReLU networks up to a softmax + LayerNorm transformer. The authors also prove no finite deterministic black-box test can certify removal, exhibiting an edit that passes exhaustive testing yet fails on an arbitrarily small survivor pocket.
- Certifies skill removal and preservation for all inputs in continuous regions
- Bound propagation handles roughly 9x the perturbation dimension of exact solvers
- Proves black-box testing alone can never certify capability removal
- Guarantees limited to small standard-architecture networks with decidable skill specifications
From the original article
1 min read at arxiv.orgMechanistic edits (ablations, weight edits, activation steering) are the standard tools for unlearning a harmful capability from a neural network while preserving useful ones. Current approaches validate their effects only by testing, which can never cover an entire continuous region of inputs.
By Md Sazid Uddin · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.