Reproduce “ASCENT: First-Order Optimal Fine-Tuning with Recalibration for Safety--Utility Co-Enhancement” 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
Weiwei Qi, Chongyu Wang, Tianhang Zheng, Zefeng Wu, Zhilin Guo, Xiaojun Jia, Zhongjie Ba, Kui Ren
ASCENT fine-tunes language models for downstream tasks while limiting safety regressions. It estimates safety-relevant weight directions from harmful prompts, preserves task-update components compatible with its safety objective, shrinks conflicting components, and periodically recalibrates. Across four models and four tasks, it improves utility and lowers JailbreakBench attack success rate in all 16 settings; these benchmark results do not guarantee broader safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.