Introduction
Recent interpretability research has popularized the idea that refusal in Large Language Models (LLMs) is mediated by a single linear direction in the model's internal residual stream. This "single direction" hypothesis suggests that by identifying and subtracting a specific "refusal vector," one can bypass safety alignments. However, this paper challenges this simplicity, arguing that refusal is a multi-dimensional phenomenon governed by complex geometric structures called "concept cones."
Methodology: Beyond Simple Linear Projections
The authors move away from standard Difference-in-Means (DIM) approaches, which typically subtract average harmful activations from harmless ones. Instead, they introduce Refusal Direction Optimization (RDO), a gradient-based method designed to find vectors that surgically mediate refusal without degrading general model capabilities.
This approach is extended to Refusal Cone Optimization (RCO), which identifies multi-dimensional polyhedral cones where every direction within the cone triggers refusal behavior. Furthermore, the authors introduce the concept of Representational Independence (RepInd). Unlike simple geometric orthogonality, RepInd identifies directions that are mechanistically distinct, ensuring that interventions in one direction do not interfere with downstream representations of another.
Key Insights: Cones and Independence
The study reveals that refusal behavior exists within multi-dimensional subspaces (up to 5D in some models), rather than along a single line. This discovery implies that there isn't just one "refusal neuron" but a whole region of activation space signaling the model to stop.
A critical finding is that orthogonality does not equal independence. In the non-linear environment of an LLM, two orthogonal vectors can still be functionally linked. The RepInd framework allows researchers to find truly separate refusal pathways. When these independent directions are ablated together, they exhibit an additive effect on jailbreak success, proving they tap into complementary internal mechanisms.
Empirical Results
The proposed RDO method demonstrates high efficacy across various model families, including Gemma 2, Qwen 2.5, and Llama 3.
As shown in the results, RDO-based directions often outperform the DIM baseline in jailbreak effectiveness. Crucially, RDO is more "surgical," resulting in significantly fewer side effects on general reasoning tasks. For instance, on the TruthfulQA benchmark, RDO maintained performance much closer to the original model compared to the DIM baseline, which often caused broader degradation. In Gemma 2 2B, ablating the top-4 representationally independent directions achieved an Attack Success Rate (ASR) of over 80%, surpassing single-direction baselines.
Conclusion
By uncovering the geometric complexity of refusal, this research moves the field beyond the "single direction" myth. Understanding that refusal is mediated by multi-dimensional cones and independent representational pathways is essential for developing more robust safety alignments and sophisticated monitoring tools for LLMs.
Summary updated