I compared three LLM customization methods—Supervised Fine‑Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning with Verifiable Rewards (RLVR)—and ran an experiment on SageMaker AI using Qwen‑3 0.6B with a numeric verifier on GSM8K problems.
SFT needs the prompt and an ideal completion. The model learns to reproduce that response.
DPO needs a preferred and rejected response for the same prompt.
RLVR needs the prompt, a reference answer, and a programmatic verifier.
The base model solved 736 of 1,319 test problems, reaching 55.80% exact‑match accuracy.
After 100 GRPO steps, the RLVR model solved 774 problems and reached 58.68%.
Training corrected 133 baseline failures but broke 95 previously correct answers. The net improvement was 38 problems, or 2.88 percentage points.