Latent Adversarial Training Improves the Representation of Refusal
This addresses the problem of understanding and improving the robustness of refusal behavior in language models for AI safety, though it is incremental as it builds on prior work on refusal directions and adversarial training.
The paper investigated how Latent Adversarial Training (LAT) reorganizes the representation of refusal behavior in language models, finding that LAT concentrates refusal in the first two SVD components, explaining approximately 75% of variance, and leads to more effective transferable refusal vectors for attacks.
Recent work has shown that language models' refusal behavior is primarily encoded in a single direction in their latent space, making it vulnerable to targeted attacks. Although Latent Adversarial Training (LAT) attempts to improve robustness by introducing noise during training, a key question remains: How does this noise-based training affect the underlying representation of refusal behavior? Understanding this encoding is crucial for evaluating LAT's effectiveness and limitations, just as the discovery of linear refusal directions revealed vulnerabilities in traditional supervised safety fine-tuning (SSFT). Through the analysis of Llama 2 7B, we examine how LAT reorganizes the refusal behavior in the model's latent space compared to SSFT and embedding space adversarial training (AT). By computing activation differences between harmful and harmless instruction pairs and applying Singular Value Decomposition (SVD), we find that LAT significantly alters the refusal representation, concentrating it in the first two SVD components which explain approximately 75 percent of the activation differences variance - significantly higher than in reference models. This concentrated representation leads to more effective and transferable refusal vectors for ablation attacks: LAT models show improved robustness when attacked with vectors from reference models but become more vulnerable to self-generated vectors compared to SSFT and AT. Our findings suggest that LAT's training perturbations enable a more comprehensive representation of refusal behavior, highlighting both its potential strengths and vulnerabilities for improving model safety.