List your company

Mathematical Reasoning on MATH500 (Training Steps to Accuracy)

160Training Steps to Target AccuracyDIFF
133.6311.8490668.2May 26, 2026

Evaluation results

MethodLinks
DIFFModel=Qwen2.5-Math-7B2026.05160
GRPOModel=Qwen2.5-Math-7B2026.05180
GAINRLModel=Qwen2.5-Math-7B2026.05180
SAERLGModel=Qwen2.5-Math-7B2026.05200
SAERLDModel=Qwen2.5-Math-1.5B2026.05340
SAERLGModel=Qwen2.5-Math-1.5B2026.05380
DAPOModel=Qwen2.5-Math-1.5B2026.05400
ADARFTModel=Qwen2.5-Math-7B2026.05400
DIFFModel=Qwen2.5-Math-1.5B2026.05440
GAINRLModel=Qwen2.5-Math-1.5B2026.05480
GRPOModel=Qwen2.5-Math-1.5B2026.05560
ADARFTModel=Qwen2.5-Math-1.5B2026.05820