Pruning "Unsafe Tickets": A Resource-Efficient Framework for Making LLMs Safer and More Robust

Researchers propose a resource-efficient LLM pruning framework that removes parameters that trigger unsafe behavior, significantly reducing unsafe generations and enhancing robustness to jailbreak attacks, with minimal loss in utility. For models that remain unsafe after alignment, such as Mistral and LLaVA, the method uses gradient-free attribution to locate unsafe subnetworks, requires only a small amount of GPU resources, and generalizes across architectures and quantized variants. The authors infer, based on the lottery ticket hypothesis, that models contain "unsafe tickets," and that pruning can reveal "safe tickets," making the approach suitable for deployment in resource-constrained settings.