model_group stringlengths 7 21 | public_exam_auto float64 45.6 88 | public_exam_human float64 42.2 85 | citation float64 17.5 85.6 | constraint float64 42.8 98.7 | argument float64 34.5 94.1 | real_case_auto float64 31.6 92.8 | real_case_human float64 40.8 87.1 | overall_auto float64 38.6 90.4 | overall_human float64 41.5 83 |
|---|---|---|---|---|---|---|---|---|---|
GPT-5.5 High | 88 | 85 | 85.6 | 98.7 | 94.1 | 92.8 | 77.9 | 90.4 | 81.5 |
GPT-5.5 Medium | 79.7 | 75.9 | 81.7 | 98.6 | 92.7 | 91 | 79.6 | 85.3 | 77.8 |
Gemini 3 Flash Low | 85 | 80.6 | 74 | 74.1 | 71.7 | 73.3 | 85.4 | 79.1 | 83 |
Qwen3.8-2.4T-A95B | 82.5 | 72.2 | 80.3 | 85.2 | 81.2 | 82.2 | 87.1 | 82.4 | 79.6 |
Gemini 3 Flash | 84.6 | 83.1 | 73.8 | 75.6 | 71.3 | 73.6 | 82.5 | 79.1 | 82.8 |
GPT-5.4 Mini | 85.6 | 75.6 | 73.6 | 90.2 | 84.9 | 82.9 | 76.7 | 84.2 | 76.1 |
Gemini-3.5-Flash | 85.2 | 79.7 | 73.3 | 74 | 72.6 | 73.3 | 81.2 | 79.3 | 80.5 |
Kimi-K3 | 76.9 | 68.1 | 77.9 | 86.1 | 80.6 | 81.6 | 77.5 | 79.2 | 72.8 |
GLM-5.3 | 75.8 | 64.7 | 79.1 | 87 | 83.2 | 83.1 | 80 | 79.5 | 72.3 |
Grok 4.5 | 74.4 | 74.1 | 74 | 89.7 | 81.3 | 81.7 | 72.9 | 78 | 73.5 |
DeepSeek-V4-Pro | 73.7 | 66.9 | 67.2 | 86.1 | 78.9 | 77.4 | 80 | 75.6 | 73.4 |
Kimi K2 Thinking | 76.6 | 69.4 | 72.9 | 75.5 | 72.1 | 73.5 | 76.7 | 75 | 73 |
DeepSeek V3.2 | 70.1 | 71.9 | 69.8 | 79.3 | 72.4 | 73.8 | 78.8 | 71.9 | 75.3 |
DeepSeek R1 | 75.8 | 71.6 | 67.5 | 79.3 | 70.3 | 72.3 | 70.8 | 74 | 71.2 |
Grok 4.3 | 74.4 | 71.9 | 59.1 | 85.3 | 70.1 | 71.5 | 70.8 | 73 | 71.4 |
Gemini 2.5 Flash | 71.9 | 62.5 | 60.3 | 84.6 | 72 | 72.3 | 75.4 | 72.1 | 69 |
Deep Cogito 671B | 67.6 | 66.9 | 63.6 | 78.8 | 67.8 | 70 | 76.3 | 68.8 | 71.6 |
GLM-5 low | 68.8 | 66.3 | 68.8 | 81.3 | 72.2 | 74.1 | 70.4 | 71.4 | 68.3 |
Qwen3 Max Thinking | 68.8 | 62.8 | 70.7 | 82.5 | 74.2 | 75.8 | 71.7 | 72.3 | 67.2 |
Grok 4.20 | 68 | 67.5 | 58.5 | 76.3 | 68.8 | 67.9 | 75.4 | 67.9 | 71.5 |
Qwen3 Max | 69.5 | 60.9 | 70.2 | 81.8 | 74.5 | 75.5 | 72.1 | 72.5 | 66.5 |
DeepSeek-V4-Flash | 69.9 | 61.6 | 67.1 | 67.6 | 68.1 | 67.6 | 72.9 | 68.8 | 67.2 |
Qwen3.8-27B | 67.7 | 58.4 | 68.2 | 80.3 | 71.6 | 73.3 | 72.5 | 70.5 | 65.5 |
Qwen3 235B A22B | 66.3 | 62.2 | 61.2 | 69.2 | 65.6 | 65.3 | 62.5 | 65.8 | 62.3 |
Command A | 62.3 | 59.7 | 52.4 | 80.2 | 64.2 | 65.6 | 68.3 | 64 | 64 |
Qwen3.5-9B | 59.1 | 45.6 | 61.9 | 69.7 | 65.3 | 65.6 | 75 | 62.4 | 60.3 |
Command R+ | 58.2 | 56.9 | 43.5 | 61.7 | 56.9 | 54 | 58.8 | 56.1 | 57.8 |
LLaMA 3.1 8B Instruct | 45.6 | 42.2 | 17.5 | 42.8 | 34.5 | 31.6 | 40.8 | 38.6 | 41.5 |
LegalScope
Measuring Exam-to-Case Transfer in LLM Legal Reasoning
AI Measurement Science (AIMS) Workshop at COLM 2026
Project page · Paper · GitHub
Hongyu Wang, Rilyn R. Han, Yilun Zhao, Xuandong Zhao, and Arman Cohan.
What is available here
This release contains workshop metadata and aggregate model results, synchronized on October 3, 2026. It replaces the earlier 868-exam / 76-case / 20-model metadata preview.
The files currently available are:
| Configuration | Contents |
|---|---|
model_performance |
28 rows of rounded model-level scores from paper Table 3 |
model_roster |
The 28 model groups evaluated in the workshop paper |
source_composition |
Counts by jurisdiction, source or legal category; these rows describe overlapping breakdowns and should not all be summed together |
data/metadata/dataset_summary.json |
Structured benchmark counts and release scope |
The full 861 exam questions, 276 case prompts, reference answers, model responses and lawyer review sheets are not included in this public release. The metadata is not an evaluation split and cannot reproduce the complete evaluation. The viewer's automatically assigned train split is a loader convention for these metadata tables, not a training set.
Benchmark described by the paper
LegalScope evaluates whether success on public legal exams transfers to closed-book reasoning about real cases.
| Component | Size |
|---|---|
| Public legal-exam questions | 861 |
| United States / China / United Kingdom / Australia | 603 / 94 / 86 / 78 |
| Real-case issue–stance prompts | 276 |
| De-identified Chinese judgments | 56 |
| Legal issues, each with support and opposition prompts | 138 |
| Model groups | 28 |
| Exam / case model responses | 24,108 / 7,728 |
| Total model responses | 31,836 |
Case categories are Tort (80 prompts), Contract (72), Criminal (54), Intellectual Property (34), Administrative (22), Civil Procedure (8), and Property (6).
Exam scoring compares answers with references on a 0–4 scale. Case scoring separates citation relevance, constraint extraction, and argument validity. Published aggregate scores are normalized to 0–100.
Human validation covers 80 exam items × 28 groups (2,240 answers) and 10 case prompts × 28 groups (280 answers). Two independent lawyers score the same 280 case answers. The 2,520 unique answers should not be double-counted.
Main results and interpretation
- Automatic exam and case scores correlate across the 28 groups (Pearson 0.817; Spearman 0.708), but rankings and variant gains do not transfer uniformly.
- Mean citation relevance is 66.9, compared with 72.6 for argument validity. Lawyer scores show the same direction descriptively, while the pooled lawyer confidence interval crosses zero.
- Answer-level automatic–human agreement is higher for exam answers (Pearson 0.910) than case answers (0.312, pooling two lawyers).
Human-score columns use the validation subsets, not the full tracks. Scores are rounded to one decimal. The different track rubrics do not equate their difficulty. Historical controls and fixed-answer stability audits use separate pools, described in the results summary.
Loading the published results
from datasets import load_dataset
results = load_dataset("Hongyu513/LegalScope", "model_performance", split="train")
roster = load_dataset("Hongyu513/LegalScope", "model_roster", split="train")
CSV and JSON files can also be downloaded directly from the Files tab. Configurations are explicit so tables with different schemas are not concatenated into rows containing unrelated null fields.
Scope, limitations and license
This is a research benchmark, not legal advice or a measure of readiness for legal practice. The paper describes de-identification and expert validation; the public files here contain no case narratives or individual lawyer notes. The model roster follows the workshop paper and excludes additional experiments outside that roster.
The existing MIT label applies to this published metadata and documentation release. It does not grant redistribution rights to third-party exam questions, judgments or reference materials that are not included here. See the data card, scoring protocol, and release notes.
Citation
@inproceedings{wang2026legalscope,
title = {{LegalScope}: Measuring Exam-to-Case Transfer in {LLM} Legal Reasoning},
author = {Wang, Hongyu and Han, Rilyn R. and Zhao, Yilun and Zhao, Xuandong and Cohan, Arman},
booktitle = {AI Measurement Science Workshop at COLM 2026},
year = {2026},
url = {https://openreview.net/forum?id=BNx62Wx1ej}
}
- Downloads last month
- 30