Shows feature steering can shift targeted behavior without hurting capabilities only within a narrow range, past which the model degrades. Informs expectations for activation-steering as a control technique.
Key takeaways · AI-distilled
Anthropic ran two social bias evaluations covering 11 bias types and two capability evaluations against every one of the 29 steered features, measuring both how targeted each feature was and what it cost.
Using MMLU accuracy as a capability proxy, the study found a steering range that left capabilities intact, and, unexpectedly, the same sweet spot held across all 29 features tested.
Steering had off-target effects: features thought to relate to gender bias also significantly shifted age bias, so a feature's firing contexts did not fully predict what steering it would do.
One neutrality feature reduced social bias across nine dimensions without much impact on the capabilities tested, the most encouraging result in what Anthropic calls mixed, preliminary findings.
articleRole Steering of Language Models for Social SimulationsIsaac Song, Mohammed Rehan Parwani, Glenn Matlin, Emile Anand, Akhil Theerthala, Arjun Chatterjee, Maria Kostylew, Yonadav G. Shavit, Sebastien Krier, Mark Riedl
articleDoes Fine-Tuning Undo Activation Steering? Behavioural Recovery Without Weight-Edit ReversalPhilipp E. Glass, Allan Tucker, Yongmin Li, Alina Miron
videoSteering LLM Behavior Without Fine-TuningHugging Face