Legal Reasoning on PRBench Legal
0.482GPT-5 Grader ScoreQwen3.5-4B RL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen3.5-4B RLTraining Data=Agentic Self-Instruct2026.06 | 0.482 | 0.436 | |
| Qwen3.5-397BRL=No RL2026.06 | 0.446 | 0.402 | |
| Qwen3.5-4B RLTraining Data=CoT Self-Instruct2026.06 | 0.422 | 0.389 | |
| Qwen3.5-4BRL=No RL2026.06 | 0.329 | 0.296 |