This project develops a hybrid reinforcement learning framework for structured reasoning on legal multiple-choice question answering.
Training design
- Designed a Zero-RL → distilled-CoT SFT → GRPO training pipeline.
- Structured outputs around statute citation and option-by-option analysis.
- Reached 57.6% accuracy, outperforming SFT-only and RL-only baselines.
Read the thesis for the training design, experimental setup, and ablation results.