Reproduce “PROSE: Preservation-Constrained Selective Editing for Guilty Bias in Legal Judgment Prediction” 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
Yimeng Wang, Zeyuan Tian, Xinyu Wang, Danyang Gao, Chenglei Shen, Zhongxiang Sun, Lei La, Weijie Yu
The authors introduce PROSE, a method for selectively editing legal judgment prediction models to correct charge-not-established cases while preserving supported verdicts. It combines training-derived contrasts with local margin signals, attenuates directions associated with supported guilt, and fits a bounded parameter update. On Qwen3-8B-Base, PROSE reaches 76.34% pairwise safe repair, above LJPIV’s 64.67% but below SFT’s 96.27%, on Chinese criminal-law data, not full judicial decision-making.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.