safety security
A Model Edit Came With a Regional Guarantee
Certified mechanistic edits prove skill removal and preservation across continuous input regions in small networks.
Summary
Certified mechanistic edits prove skill removal and preservation across continuous input regions in small networks.
The method certifies that disabling a circuit removes one specified skill while preserving another for every input in a region. Demonstrations range from toy ReLU networks to a standard softmax-plus-LayerNorm transformer, with sound bound propagation reaching about nine times the perturbation dimension of an exact solver. The guarantees remain limited to small models and decidable skill specifications.
Why it matters
Certified mechanistic edits prove skill removal and preservation across continuous input regions in small networks.
Limits and context
- The guarantees remain limited to small models and decidable skill specifications.
Key claims
Certified mechanistic edits prove skill removal and preservation across continuous input regions in small networks.
Qualification: The guarantees remain limited to small models and decidable skill specifications.
Evidence: source-2026-10-05-014
Sources
- arXiv preprint 2610.03502arXiv · primary research
Corrections
No corrections have been recorded for this story.