Reproduce “Restoring General and Improving Domain-Specific Safety in Fine-Tuned LLMs Without Retraining” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vishnu Kabir Chhabra, Ding Zhu, Zhihui Zhu, Xueru Zhang, Mohammad Mahdi Khalili
SPECTRA repairs refusal weakening after domain fine-tuning through a training-free weight edit. It transfers the base model’s refusal direction, gated by a low-rank subspace estimated from harmful-prompt activations, without gradient updates or inference-time steering. Across tested Llama-2 and Gemma2 variants, it lowers attack success; Gemma2 medical GCG success falls from 36% to 0%, though exceptions and untested attacks limit the findings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.