Optimizing Against Safety Representations: Activation-Guided Adversarial Suffixes and the Geometry of Refusal

Recent research examines the geometry of representations in language models. Studies on Gemma-2-2B, Gemma-2-9B, and Qwen3-4B reveal that 1D manifolds with place-cell feature tiling emerge for locally computable ordinal tasks, while complex tasks produce higher-dimensional representations. Simultaneously, another study explores behavioral alignment and safety representations. By introducing Activation-Guided GCG, researchers use adversarial suffix attacks to target internal refusal directions, discovering that suppressing refusal globally across all layers and positions is more effective than localized targeting.