LegalCiteBench - Misleading Answer Rate: leaderboard

LegalCiteBench (2026) tests whether LLMs, without retrieval, can recover and verify the case authorities cited in 1,000 U.S. judicial opinions from the Case Law Access Project. The misleading answer rate is the share of low-scoring retrieval and matching responses that still assert a concrete citation or case rather than abstaining, in percent; lower is better.

Metric: Misleading Answer Rate (%): share of low-scoring responses (task score at most 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 2037. 21 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5 Mini89.24#364
2DeepSeek-R1-Distill-Qwen-7B94.21#8055
3DeepSeek R1 Distill Qwen 1.5B94.62#9215
4Claude Sonnet 4.596.79#261
5O4 Mini96.85#368
6Qwen 3 14B96.97#1269
7Qwen 2.5 7B97.53#2929
8DeepSeek V3.197.58#511
9SaulLM-54B97.69
10Gemini 2.5 Flash98.21#463
11Claude Haiku 4.598.31#583
12Qwen 3 30B A3B98.35#1156
13Llama 3.1 70B98.47#1499
14Mistral 7B98.69#7794
15Qwen 3 4B99.01#2348

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=legalcitebench-misleading-answer-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.