RL vs. SFT for Mathematical Reasoning in LLMs
Implemented GMPO and its evaluation pipeline for a compute-controlled comparison of PPO, GRPO, GMPO, RLOO, and SFT on Qwen3-8B.
In an MIT 6.4610 team project, we ran a compute-controlled comparison of PPO, GRPO, GMPO, and RLOO against supervised fine-tuning on Qwen3-8B. My contribution was implementing GMPO and its evaluation pipeline.
GMPO reached 74.2% on GSM8K versus 76.7% for supervised fine-tuning. The named branch and filtered commit history provide direct evidence of my implementation work.