Legal Reasoning on LegalBench CUAD Cardlytics Buffalo Wild Wings PF Hospitality 2023
82.7Accuracy (Cardl)Agentic Adversarial QA
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Agentic Adversarial QATokens=96k, Backbone=LLaMA3-8b, Evaluation Protocol=few-shot2026.02 | 82.7 | 79.6 | 85.7 | 82.7 | |
| EntigraphTokens=6.7M, Backbone=LLaMA3-8b, Evaluation Protocol=few-shot2026.02 | 80.4 | 76.5 | 82 | 79.6 | |
| Knowledge-InstructTokens=159k, Backbone=LLaMA3-8b, Evaluation Protocol=few-shot2026.02 | 78.6 | 70.4 | 75.8 | 75 | |
| Model-indep. QATokens=147k, Backbone=LLaMA3-8b, Evaluation Protocol=few-shot2026.02 | 75 | 74.1 | 78.3 | 75.8 | |
| Paraphrase ×6Tokens=149k, Backbone=LLaMA3-8b, Evaluation Protocol=few-shot2026.02 | 68.5 | 70.4 | 77 | 71.9 | |
| No Extra DataTokens=0, Backbone=LLaMA3-8b, Evaluation Protocol=few-shot2026.02 | 67.3 | 69.1 | 72.1 | 69.5 |