There is more than one SOM direction: multi-directional refusal suppression in language models

The paper proposes using self-organizing maps (SOM) to extract multiple refusal directions from inside a language model, and to suppress refusal by ablating these directions. The authors first show that SOM can generalize the previous approach of computing a single refusal direction from the difference in centroids of harmful and harmless prompt representations, then train a SOM on harmful prompt representations to identify multiple neurons, subtracting the harmless representation centroid from each neuron to obtain a set of directions expressing the refusal concept. In a larger-scale experimental setting, ablating multiple directions outperforms not only the single-direction baseline but also dedicated jailbreak algorithms, thereby achieving effective refusal suppression. The paper concludes by analyzing the mechanistic implications of the method.