IC-590Tuning only the identified safety neurons (SN-Tune) reduces harmful scores of instruction-tuned and base models by over 90 points while preserving general capability.

Yiran Zhao, Wenxuan Zhang, Yuxi Xie, Anirudh Goyal, Kenji Kawaguchi, Michael Shieh

SourceUnderstanding and Enhancing Safety Mechanisms of LLMs via Safety-Specific Neuron

The authors fine-tune only the identified safety neurons (setting gradients of all other parameters to zero) on 50 safety documents for 1 epoch at learning rate 1e-6. For instruction-tuned models, this reduces average harmful scores from 93.5 to 3.0 (Vicuna-13b-v1.5), 65.5 to 2.0 (Llama3-8b-instruction), and 70.8 to 4.5 (Mistral-7b-instruct-v0.2). For base models, it reduces scores from 98.8 to 5.3 (Llama2-7b-base), 100.0 to 13.5 (Llama3-8b-base), and 100.0 to 13.8 (Mistral-7b-v0.1). General capability (MMLU, ARC-C, GSM8K) is largely preserved or slightly improved. Training takes only 2-4 minutes versus 23-43 minutes for the Circ-Break baseline.

Evidence
interventional
Key metric
harmful score 93.5→3.0 (Vicuna-13b-v1.5), 65.5→2.0 (Llama3-8b-instruction), 70.8→4.5 (Mistral-7b-instruct-v0.2); 98.8→5.3 (Llama2-7b-base), 100.0→13.5 (Llama3-8b-base), 100.0→13.8 (Mistral-7b-v0.1); training cost 2-4 min vs 23-43 min for circ-break
Caveat
Evaluation uses attack success rate on Harmful Behavior testset with 4 attack methods; the safety corpus is only 50 documents from Zou et al. (2024).
Model
Vicuna Vicuna-13b-v1.5, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 Mistral 7B Instruct v0.2, Mistral 7B v0.1, Llama 2 / Llama 2 base Llama2-7b-base, Llama 3 8B Base
Datasets
MMLU / MMLU-Math [eval], ARC-Challenge [eval], GSM8K [eval]
Related findings
IC-591
Extraction
automatic-extraction