🧪 Test?View on arXiv
Feature Elicitation for SAE Steering of Deceptive Behaviour in Small LLMs
Author1, Author2, Author3, Author4, Author5
deceptive alignmentsparse autoencodersalignment fakingmechanistic intervention
2610.20638
Builder Relevance
1h ago70%
Abstract
This research investigates the mitigation of deceptive alignment in small LLMs through SAE steering features associated with deontological ethics.
Reality Card
Core Claim
The study demonstrates that steering SAE features can significantly reduce the compliance gap in the Gemma-3-27B-IT model, marking the first evidence of a mechanistic intervention's effectiveness in this context.
Method / Result
The compliance gap was reduced to non-significance in Gemma-3-27B-IT.
Limitations
The tested configurations were ineffective in LLaMA-3-8B-Instruct, raising concerns about generalizability.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.
No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.