Reproduce “TERM: Target-Free Jailbreak Defense via Entropy-Guided Residual Modulation” 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
Yunkai Dang, Wanyu Zhang, Meiyi Zhu, Qi Fan, Tianyu Ding, Wenbin Li
TERM adapts an inference-time intervention to changes in a model’s attention and residual updates, without using new weights, an external safety classifier, or a predefined steering direction. On LLaMA-3.1-8B, it averages 97.00% defense success across seven jailbreak attacks and 1.43% over-refusal across two benign benchmarks, versus 91.86% and 2.13% for AdaSteer. These results are limited to tested models and benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.