Legal Reasoning on PRBench Legal-Hard
0.377GPT-5 Grader ScoreQwen3.5-4B RL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen3.5-4B RLTraining Data=Agentic Self-Instruct2026.06 | 0.377 | 0.331 | |
| Qwen3.5-397BRL=No RL2026.06 | 0.341 | 0.294 | |
| Qwen3.5-4B RLTraining Data=CoT Self-Instruct2026.06 | 0.32 | 0.3 | |
| Qwen3.5-4BRL=No RL2026.06 | 0.241 | 0.219 |