Is this Citation on Point?
LLMs can spot fake legal cases, but still miss citations that point to the wrong page of a real case.
The paper tests proposition-level citation support using altered citations from real legal corpora. Fourteen model setups caught 93-100% of wrong-case substitutions, but performed much worse when only the pinpoint page was changed. On court opinions, wrong-pinpoint detection ranged from 37-61%; on briefs, 52-83%. Even GPT-5.4 with high reasoning missed 40% of court-opinion pinpoint mismatches and 18% in briefs. HF Daily Papers' note
The paper tests proposition-level citation support using altered citations from real legal corpora. Fourteen model setups caught 93-100% of wrong-case substitutions, but performed much worse when only the pinpoint page was changed. On court opinions, wrong-pinpoint detection ranged from 37-61%; on briefs, 52-83%. Even GPT-5.4 with high reasoning missed 40% of court-opinion pinpoint mismatches and 18% in briefs. HF Daily Papers' note
score 4