DLawBench - Fidelity: leaderboard
DLawBench (2026) has LLMs act as lawyers in multi-turn consultations with simulated clients (cooperative, dependent, withdrawn or adversarial) over real Chinese-law and U.S.-law cases, then write a legal memo that a three-judge LLM panel scores against expert rubrics. This board is fidelity, the share of claims in the memo that the consultation actually supports. Scores are rates on a 0 to 100 scale; higher is better.
Metric: Fidelity score (0-100), a rate scaled by 100: share of factual and inferential claims in the legal memo that the consultation supports, averaged over the two jurisdictions, on 461 real Chinese-law and U.S.-law cases replayed as multi-turn consultations with a Claude Sonnet 4.6 client simulator in four narrative styles; judged by a GPT-5.1, Claude Opus 4.6 and Gemini 3.1 Pro panel (median; same-vendor judges recuse), empty memos count as zero; provider decoding defaults; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 94 | #123 |
| 2 | GPT-5.2 | 93.6 | #176 |
| 3 | GPT-5.5 | 93.4 | #49 |
| 4 | DeepSeek V3.2 (Thinking) | 89.8 | #406 (DeepSeek V3.2) |
| 5 | Claude Sonnet 4.6 | 88.9 | #161 |
| 6 | GLM-5.1 | 88.4 | #179 |
| 7 | GLM-5 | 88 | #260 |
| 8 | Seed 2.0 Pro | 87.7 | #211 |
| 9 | Claude Opus 4.6 | 87.6 | #102 |
| 10 | Claude Opus 4.7 | 87.3 | #79 |
| 11 | Qwen 3.6 Max Preview | 87 | #159 |
| 12 | Kimi K2.6 | 86.8 | #178 |
| 13 | Qwen 3.6 Plus | 86.5 | #222 |
| 14 | Qwen 3.5 Plus | 85.9 | #232 |
| 15 | DeepSeek V4 Pro | 85.1 | #188 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dlawbench-fidelity · How It Works · Data refreshed daily, snapshot 2026-11-13.