Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs
Open paper
Benchmark
Model
Rank
Results
math-word-problem-solving-on-math
Qwen2-72B-Instruct-Step-DPO (0-shot CoT, w/o code)
#32
Accuracy: 70.8
arithmetic-reasoning-on-gsm8k
Qwen2-72B-Instruct-Step-DPO (0-shot CoT)
#35
Accuracy: 94.0
Rank counts only results with a code link.