**Original paper:** [https://arxiv.org/abs/2608.12571v1](https://arxiv.org/abs/2608.12571v1)

# Summary: 2026-08-12_20-28-55Z_IsthisCitationonPoint.md
Saved: 2026-08-13 22:26
Source: 2026-08-12_20-28-55Z_IsthisCitationonPoint.md
Model: None

---

## Summary  
The paper investigates whether legal citations generated by LLMs actually support the propositions they are used to argue for, beyond simple factual correctness. It introduces a controlled perturbation task that replaces either the cited case or only the pinpoint page within a case, then evaluates 14 model configurations on this data. The study reveals that models excel at detecting wrong‑case errors but struggle with wrong‑pinpoint errors, often accepting citations based on topical overlap rather than precise support. Prompting for page‑level verification improves recall while increasing false positives, highlighting a fundamental conflation of topic relevance and citation justification.  

## Key Contributions  
- Finding 1: Models achieve 93‑100% detection of wrong‑case corruptions but only 37‑61% on court opinions and 52‑83% on briefs for pinpoint mismatches.  
- Finding 2: High reasoning effort (e.g., GPT‑5.4) reduces errors to ~40% on opinions and 18% on briefs, narrowing but not closing the gap.  
- Finding 3: Prompting models to verify support at the cited page raises recall while increasing false positives, indicating a trade‑off between sensitivity and specificity.  

## Methodology  
The authors constructed a dataset by taking real legal citations from two corpora and applying controlled perturbations—either swapping the entire case or altering only the pinpoint page. They generated 14 model configurations (different prompting strategies and reasoning prompts) and evaluated them on both opinion pages and briefs, measuring recall and false‑positive rates.  

## Results  
Wrong‑case corruptions are caught at high precision; wrong‑pinpoint errors remain problematic, with detection rates varying widely by document type. The best models still miss a substantial fraction of pinpoint mismatches despite advanced prompting or reasoning.  

## Significance  
This work uncovers a critical blind spot in LLM evaluation for legal use, showing that citation support verification is distinct from topic alignment and that current systems often misattribute relevance to factual correctness, which could lead to erroneous legal arguments.  

## Related Concepts  
Legal citation, proposition‑level support, pinpoint citation, topical overlap, hallucinated citations, LLM reasoning, false positive/recall trade‑off, legal briefs vs court opinions.