AWhat changed
Trained representation-level adapters to redirect activations associated with unwanted behavior while retaining representations for benign behavior, testing a safeguard designed to interrupt problematic generation pathways.
Scientific scope
Trained representation-level adapters to redirect activations associated with unwanted behavior while retaining representations for benign behavior, testing a safeguard designed to interrupt problematic generation pathways. The method requires model internals, adapter training and harmful/retain datasets. Reported tests focus on selected single-turn harmful-output attacks. Cygnet combines rerouting with other representation controls, so its complete gains cannot be assigned to rerouting alone. No universal robustness or cost-free filter is proved.
Evidence
The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The method requires model internals, adapter training and harmful/retain datasets. Reported tests focus on selected single-turn harmful-output attacks. Cygnet combines rerouting with other representation controls, so its complete gains cannot be assigned to rerouting alone. No universal robustness or cost-free filter is proved.
Contributors
- The named authors in the research record · SCIENCE: Source-listed authors for the exact contribution; experimental subjects are not autonomous scientific authors. Source support and predecessor work retain separate roles. Individual effort shares and unresolved identities are not inferred.
Provisional WRS assessment, edition 2026-10-23.4. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction