Reproduce “Cited but Not Consulted: A Counterfactual Audit of Legal Chain-of-Thought Faithfulness” 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
Saisab Sadhu, Shreeyans Arora, Pratinav Seth
The authors test whether legal-model verdicts depend on named authorities by holding case facts fixed and replacing the target authority. Across seven models and four tasks, authority naming was 66.7%–100%, while verdict-swap sensitivity varied. In ECHR Pair A, Llama-3.1-8B named the authority in 100% of cases but changed verdict in 58.6%; prompted behavior does not establish internal causal use.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.