CanLegalRAGBench (Oracle Retrieval) - Answer Accuracy: leaderboard
Legal question answering over Canadian case law: 532 realistic lay and legal-expert questions with answers written by legal annotators; CanLegalRAGBench (Zhao et al., 2026). The generator answers from chunks retrieved only from each question's ground-truth case documents (one fixed retriever for every generator), and a Gemini 2.5 Pro evaluator splits the answer into atomic claims; the score is the share of those claims entailed by the expert answer, in percent (526 of the 532 questions for Qwen). Higher is better.
Metric: Claim accuracy (%; share of the answer's atomic claims entailed by the expert answer's claims, Gemini 2.5 Pro FActScore-style evaluator; 532 queries, 526 for Qwen; retrieval confined to the ground-truth case documents). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B (Thinking) | 76.46 | #803 (Qwen 3.5 9B) |
| 2 | Gemini 2.5 Flash | 66.93 | #452 |
| 3 | Gemma 4 E4B | 63.51 | #2217 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=canlegalragbench-oracle-retrieval-answer-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.