Conference
Open access
Sep 2026
Detoxifying Large Language Models via Localized Feature Editing with Sparse Autoencoders
L DeLFE, a lightweight controlled detoxification approach that identifies specific toxic features across model layers and performs targeted interventions on them, and designs three feature-level strategies that adjust intervention timing and strength to reconstruct the target model’s original activations.
Yu-Hu Shang, Xiang Cheng, Dian-Yun Wang et al.
· Proceedings of the Thirty-Fi... · 0 citations