Curvature-Guided Module Localization for Low-Rank Detoxification of Backdoored Large Language Models

Name: Curvature-Guided Module Localization for Low-Rank Detoxification of Backdoored Large Language Models
Uploaded: 2026-10-08T11:41:19.900311+00:00
Duration: 296 s
Description: The framework localizes modules responsible for backdoor triggers using activation patching and Fisher/K-FAC curvature analysis to find influential components. It then applies targeted low-rank parameter repairs only to these key modules, effectively neutralizing malicious behavior while preserving the model's general language capabilities.

The framework localizes modules responsible for backdoor triggers using activation patching and Fisher/K-FAC curvature analysis to find influential components. It then applies targeted low-rank parameter repairs only to these key modules, effectively neutralizing malicious behavior while preserving the model's general language capabilities.

Open the full episode