Technical Summary: LegalCiteBench

Problem Statement

Large Language Models (LLMs) are increasingly integrated into legal workflows, yet their reliability in generating specific case authorities without external retrieval (closed-book setting) remains unquantified. In common-law systems, legal arguments rely on precise citations to judicial precedents; incorrect or fabricated authorities can lead to professional sanctions and undermine trust in AI tools. While existing legal benchmarks focus on statutory reasoning, contract understanding, or general legal question answering, they do not systematically evaluate a critical failure mode: the generation of plausible-looking but incorrect citations when models are asked to provide case authorities without external grounding. The core problem is whether current LLMs can recover exact judicial citations, abstain when uncertain, or if they instead fabricate convincing but erroneous authorities.

Methodology

Dataset Construction: LegalCiteBench

The authors introduce LegalCiteBench, a benchmark comprising approximately 24,000 evaluation instances derived from 1,000 real U.S. judicial opinions sourced from the Case Law Access Project (Harvard Law School). The construction pipeline involves:

1. Source Material: Sampling 1,000 recent root cases across U.S. jurisdictions.
2. Intermediate Summarization: Using three commercial LLMs (GPT-4o-mini, Gemini-2.5-Flash, Claude-3.5-Haiku) to generate structured construction notes summarizing case backgrounds, procedural history, and legal issues. These notes are used only for prompt construction, not as ground truth.
3. Ground Truth Anchoring: Ground truth labels are strictly derived from non-generated sources: extracted citation lists from the source opinions and root-case metadata.
4. Human Validation: A targeted audit of 100 sampled root cases by three reviewers to verify construction fidelity, anonymization quality, and ground-truth correctness.

Task Taxonomy

The benchmark evaluates five citation-centric tasks divided into two levels:

• Citation-Level Tasks:
  • Cat1 (Citation Retrieval): Given a lawyer-style research question derived from a real case, the model must recover the specific authorities cited in the source opinion.
  • Cat2 (Citation Completion): Given a partial set of citations, the model must identify the remaining authorities from the source opinion.
  • Cat3 (Citation Error Detection): Given a legal analysis paragraph, the model must detect and correct citation errors (e.g., wrong volume, page, or reporter).
• Case-Level Tasks:
  • Cat4-1 (Case Matching): Given an anonymized legal scenario, the model must identify the underlying source case.
  • Cat4-2 (Case Verification and Correction): Given a scenario with an explicit citation, the model must verify its correctness and provide the correct case if the citation is wrong.

Evaluation Protocol

• Models Evaluated: 21 LLMs, including proprietary models (e.g., GPT-4o-mini, Claude-3.5-Haiku, Gemini-2.5-Flash) and open-source models (e.g., Llama 3, Qwen, DeepSeek, SaulLM).
• Judging Mechanism: An "LLM-as-a-judge" protocol using GPT-4o-mini (with specific rubrics) to evaluate outputs. Scores are normalized to a 0–100 scale.
• Metrics:
  • Citation F1: For retrieval tasks, using case-insensitive substring matching to compare predicted citations against ground truth.
  • Misleading Answer Rate (MAR): A safety metric for retrieval-heavy tasks (Cat1, Cat2, Cat4-1). It measures the proportion of low-scoring responses where the model provides a concrete citation or case answer rather than abstaining. High MAR indicates a tendency to hallucinate specific authorities rather than express uncertainty.
  • Abstention Rate: The frequency with which models explicitly state they cannot verify or identify the requested authority.

Key Results

1. Exact Citation Recovery is Universally Challenging

In the closed-book setting, exact recovery of opinion-anchored authorities is extremely difficult for all evaluated models.

• Low Scores: Even the strongest model, claude-sonnet-4.5, scored only 6.80 on Cat1 and 6.35 on Cat2. No model exceeded 7/100 on citation retrieval or completion.
• Asymmetry: There is a sharp performance gap between generation and verification. Models score significantly higher on verification tasks (Cat3: up to 75.59; Cat4-2: up to 96.07) compared to generation tasks. This suggests models can often judge the plausibility of a provided citation but cannot reliably retrieve the correct authority from memory.

2. High Misleading Answer Rates (MAR)

Models frequently provide concrete but unreliable answers rather than abstaining.

• MAR Dominance: For 20 out of 21 models, the MAR on retrieval-heavy tasks exceeded 94%. Even the best-performing model (gpt-5-mini) had an MAR of 89.24%.
• Implication: When models fail, they tend to fabricate specific authorities rather than admit uncertainty, posing a significant safety risk in legal workflows.

3. Limited Impact of Scale and Domain Pretraining

• Scaling: Increasing model size yields only marginal gains. For example, Llama-3.1-70B (3.82 on Cat1) performed only slightly better than Llama-3.1-8B (1.47 on Cat1), despite being nearly 9x larger.
• Domain Specialization: SaulLM-54B, a legal-domain model, achieved the highest score on citation error detection (Cat3: 75.59) but remained weak on retrieval (Cat1: 3.77; Cat2: 4.06) with a high MAR (97.69%). This indicates that legal-domain pretraining improves verification skills but does not resolve the grounding gap for closed-book citation retrieval.

4. Prompt-Only Mitigation

A diagnostic experiment tested whether explicit abstention instructions could reduce MAR.

• Results: Explicit instructions reduced MAR for some models (e.g., Llama-3.1-8B-Instruct reduced MAR from 100% to 62.7%), primarily by increasing the abstention rate.
• Limitation: The reduction in MAR did not correlate with improved citation correctness (F1 scores remained near zero). Prompting makes models less likely to fabricate but does not equip them with the knowledge to generate correct authorities.

Significance and Claims

The paper positions LegalCiteBench as a diagnostic framework rather than a certification of legal competence. Its primary contributions are:

1. Benchmarking Authority Generation: It establishes a standardized, auditable method for evaluating closed-book citation recovery, a critical failure mode in legal AI that existing benchmarks overlook.
2. Revealing Calibration Gaps: The results demonstrate a substantial gap in model calibration; current LLMs are prone to confident fabrication of legal authorities when external grounding is absent.
3. Safety-Oriented Evaluation: By introducing the Misleading Answer Rate (MAR), the work highlights that high accuracy on verification tasks does not guarantee safety in retrieval tasks.
4. Limitations of Current Approaches: The study concludes that neither scaling nor domain-specific pretraining currently solves the problem of exact citation recovery in closed-book settings. It argues that reliable legal AI requires retrieval-augmented systems, verification modules, or post-training strategies specifically targeting calibrated abstention, rather than relying on parametric memory alone.

The authors emphasize that these findings should not be interpreted as a failure of retrieval-augmented legal systems, but rather as a stress test revealing the risks of ungrounded generation. They advocate for independent verification of any citation produced without external grounding before use in professional legal analysis.