code generation research on inference scaling

🚀 Improvements Made

Component	Description
Advantage	Improved estimation method for more stable policy gradients.
Reward	Reward function reshaping to enhance learning signals.
Objective Function	Modified loss function to better align with final task rewards.

---

🧪 Experiments and Ablation Study

We conduct the following experiments:

Experiment Branch	Description
original_grpo	Original GRPO algorithm (baseline).
improved_advantage	Only the advantage estimation improved.
improved_reward	Only the reward function improved.
improved_objective	Only the objective function improved.
full_improvement	All improvements applied together.

Results are saved under the results/ directory, categorized by experiment.

---

📜 How to Run

1. Install dependencies:

pip install -r requirements.txt