Safe steering or falling off a cliff? Re-examining the specificity and robustness of inference-time interventions

AI reading guideNavita Goyal and Hal Daumé propose a three-dimensional framework for evaluating the specificity of model steering, and systematically evaluate five steering methods in two safety scenarios: over-refusal and faithfulness hallucination. The framework divides specificity into three dimensions: general, control, and robustness, measuring respectively whether steering maintains fluency and unrelated capabilities, whether it maintains relevant control properties, and whether it maintains these properties under distribution shift. Results show that steering effectively reduces over-refusal and largely maintains general and control specificity, but jailbreak robustness consistently decreases on Llama-3.1-8B-Instruct, Llama-3.2-3B-Instruct, Qwen-2.5-7B-Instruct, and Gemma-2-2B-it, with Llama-8B-Instruct dropping by 35% to 55%. The authors state this is the first systematic evaluation of steering specificity, arguing that effectiveness and standard specificity checks alone are insufficient to determine whether steering is safe.