LexRubric - Ethics and Safety: leaderboard
LexRubric (2026) scores open-ended answers to 649 Chinese legal questions from everyday legal consultation and judicial examination against expert-written atomic rubric items, each checked by an LLM judge. This board counts only the rubric items of the ethics and safety dimension. The score rate is earned rubric points (negative items subtract) over the maximum positive score in percent, averaged over instances; higher is better.
Metric: Dimension score rate (%): points earned from expert-written atomic rubric items of the ethics and safety dimension (weighted -10 to +10; negative items subtract) divided by the maximum positive score, averaged over instances; a Qwen3.6-27B judge checks each item, temperature 0.6, 16k output tokens; can fall below 0; higher is better (instances without a positive item in the dimension are left out). Source: arxiv.org. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 72.74 | #161 |
| 2 | Qwen 3 Max | 69.64 | #403 |
| 3 | DeepSeek V3.2 | 61.16 | #406 |
| 4 | Qwen 3.6 Max Preview | 59.82 | #159 |
| 5 | Qwen 3.5 397B A17B | 56.67 | #246 |
| 6 | DeepSeek V4 Flash | 54.72 | #258 |
| 7 | Kimi K2.6 | 42.55 | #178 |
| 8 | Kimi K2.5 | 40.86 | #256 |
| 9 | DeepSeek R1 | 40.64 | #473 |
| 10 | GPT-5.2 | 34.7 | #176 |
| 11 | DeepSeek V4 Pro | 27.59 | #188 |
| 12 | GLM-5.1 | 26.84 | #179 |
| 13 | GLM-5 | 22.03 | #260 |
Interactive version: theaggregate.ai/benchmark?slug=lexrubric-ethics-and-safety · How It Works · Data refreshed daily, snapshot 2026-11-13.