LLM Forensics: Where Do Backdoors Hide? Localizing and Controlling Trigger Mechanisms with Sparse Autoencoders
These token-trigger mechanisms decompose into distinct SAE feature directions, with separate features for trigger detection, residual-stream propagation, and later language tracking, with this role-level decomposition most likely to transfer to other trigger-based backdoors, even when the payload, layers, or circuit lo...
Wissam Antoun, Francis Kulumba, Théo Lasnier et al.
· 1 citation