Public articles linked to the same research event.
arXiv The work upgrades mechanistic edits (ablations, weight edits, activation steering) from test-based validation to behavioral certification: for every input in a continuous embedding-space region it proves that disabling a circuit removes one skill and preserves another, demonstrating removal and preservation from toy ReLU networks up to a standard softmax + LayerNorm transformer, and using sound bound propagation to reach roughly 9x the input-perturbation dimension an exact solver can handle; it further proves that no finite deterministic black-box test can certify removal, exhibiting an edit that passes exhaustive testing yet provably fails on a survivor pocket that can be made arbitrarily small.
The work upgrades mechanistic edits (ablations, weight edits, activation steering) from test-based validation to behavioral certification: for every input in a continuous embedding-space region it proves that disabling a circuit removes one skill and preserves another, demonstrating removal and preservation from toy ReLU networks up to a standard softmax + LayerNorm transformer, and using sound bound propagation to reach roughly 9x the input-perturbation dimension an exact solver can handle; it further proves that no finite deterministic black-box test can certify removal, exhibiting an edit that passes exhaustive testing yet provably fails on a survivor pocket that can be made arbitrarily small.
The work upgrades mechanistic edits (ablations, weight edits, activation steering) from test-based validation to behavioral certification: for every input in a continuous embedding-space region it proves that disabling a circuit removes one skill and preserves another, demonstrating removal and preservation from toy ReLU networks up to a standard softmax + LayerNorm transformer, and using sound bound propagation to reach roughly 9x the input-perturbation dimension an exact solver can handle; it further proves that no finite deterministic black-box test can certify removal, exhibiting an edit that passes exhaustive testing yet provably fails on a survivor pocket that can be made arbitrarily small.
The work upgrades mechanistic edits (ablations, weight edits, activation steering) from test-based validation to behavioral certification: for every input in a continuous embedding-space region it proves that disabling a circuit removes one skill and preserves another, demonstrating removal and preservation from toy ReLU networks up to a standard softmax + LayerNorm transformer, and using sound bound propagation to reach roughly 9x the input-perturbation dimension an exact solver can handle; it further proves that no finite deterministic black-box test can certify removal, exhibiting an edit that passes exhaustive testing yet provably fails on a survivor pocket that can be made arbitrarily small.