IC-506Restoring the refusal feature in Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.2, and Gemma-7B-IT causally disables all four tested adversarial attacks

Lei Yu, Virginie Do, Karen Hambardzumyan, Nicola Cancedda

SourceRobust LLM safeguarding via refusal feature adversarial training

The authors take successful adversarial prompts and, during inference, patch the residual stream activations to restore the refusal feature to its mean value on the original (unmodified) harmful inputs. This intervention dramatically reduces the attack success rate across all four attack methods and all three models, as judged by the official HarmBench classifier. General capability (MMLU, MT-Bench) is unaffected by the clamping, confirming the effect is specific to the safety decision rather than a degradation of generation quality.

Evidence
interventional
Key metric
MMLU after RF clamping: Llama-3-8B 65.9→65.5, Mistral-7B 58.9→58.1, Gemma-7B 52.0→51.8; MT-Bench: 7.24→7.07, 7.02→6.97, 6.60→6.74 (Table 5)
Caveat
The specific ASR values before and after restoration are shown only in Figure 4 (bar chart); no per-attack numeric ASR is printed in the text. The restoration uses the mean activation over the original harmful prompts, which is an approximation of the pre-attack state.
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], MMLU / MMLU-Math [eval], MT-Bench [eval]
Methods
GCG [eval], PAIR [eval], AutoDan [eval], HumanJailbreaks [eval]
Related work
Arditi et al. 2024 [builds-on]
Related findings
IC-505, IC-507
Extraction
automatic-extraction