IC-505Adversarial attacks on Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.2, and Gemma-7B-IT shift hidden representations along the negative refusal feature direction

Lei Yu, Virginie Do, Karen Hambardzumyan, Nicola Cancedda

SourceRobust LLM safeguarding via refusal feature adversarial training

The paper computes the mean representational shift induced by four distinct attack methods (GCG, PAIR, AutoDan, HumanJailbreaks) on 400 harmful prompts from HarmBench and measures the cosine similarity of that shift with the negative refusal feature vector at every layer. Across all three models and all four attacks, the shifts align significantly above chance with the refusal feature direction, indicating a shared geometric mechanism. PCA projections of the 16th-layer activations confirm that adversarially modified prompts are pushed toward the harmless cluster along the refusal axis.

Evidence
correlational
Caveat
The cosine similarity values are reported only in Figure 2 (layer-wise plots with 99% confidence intervals); no single aggregate number is printed in the text. The alignment is measured on the subset of prompts that successfully jailbreak the model, which may overstate the effect.
Model
Llama 3 8B Instruct, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 Mistral 7B Instruct v0.2, Gemma Gemma-7B-IT
Concepts
Linear representation
Datasets
HarmBench / HarmBench Prompt / HarmBench Response / HarmBench-adv [eval], AdvBench / AdvBench-50 [source]
Methods
GCG [eval], PAIR [eval], AutoDan [eval], HumanJailbreaks [eval]
Related work
Arditi et al. 2024 [builds-on], Zou et al. 2023b [context], Chao et al. 2023 [context]
Related findings
IC-506, IC-507
Extraction
automatic-extraction