SPECTRA: A training-free weight-level method to restore refusal behavior in fine-tuned models

1. 06:24
   arXiv: Dangerous Capabilities & Safety Evals

   Restoring general safety and improving domain-specific safety of fine-tuned large language models without retraining

   Researchers propose SPECTRA, a training-free weight-level method that can restore a model's refusal behavior after domain fine-tuning. The method uses activation-aware low-rank updates to transfer the base model's refusal direction into the fine-tuned model's weights, requiring neither gradient-based optimization nor domain-specific safety data. The low-rank design confines the intervention to the subspace most causally responsible for harmful inputs, making benign behavior largely unaffected by construction. The authors say that in comprehensive safety evaluations, the method markedly improves refusal behavior while maintaining general coherence and domain knowledge, and the change in false refusal rate is negligible.

Current attention 30·Peak in the comparable period 35 (Nov 10 20:00)·Change over the comparable past 24 hours -1%