LegalCiteBench - Misleading Answer Rate: leaderboard
Metric: Misleading Answer Rate (%): share of low-scoring responses (task score under 40) on Cat1, Cat2 and Cat4-1 that still give a concrete citation or case instead of abstaining, closed-book (no retrieval) on instances derived from 1,000 U.S. judicial opinions of the Case Law Access Project, greedy decoding at temperature 0 with at most 400 generated tokens, GPT-4o-mini judge; lower is better. Source: arxiv.org. Saturation forecast: Around 2040. 21 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 Mini | 89.24 |
| 2 | DeepSeek-R1-Distill-Qwen-7B | 94.21 |
| 3 | DeepSeek R1 Distill Qwen 1.5B | 94.62 |
| 4 | Claude Sonnet 4.5 | 96.79 |
| 5 | O4 Mini | 96.85 |
| 6 | Qwen 3 14B | 96.97 |
| 7 | Qwen 2.5 7B | 97.53 |
| 8 | DeepSeek V3.1 | 97.58 |
| 9 | Gemini 2.5 Flash | 98.21 |
| 10 | Claude Haiku 4.5 | 98.31 |
| 11 | Qwen 3 30B A3B | 98.35 |
| 12 | Llama 3.1 70B | 98.47 |
| 13 | Mistral 7B | 98.69 |
| 14 | Qwen 3 4B | 99.01 |
| 15 | Phi-4 | 99.1 |
Interactive version: theaggregate.ai/benchmark?slug=legalcitebench-misleading-answer-rate · How It Works · Data refreshed daily, snapshot 2026-10-07.