List your company
ResearchFormerly WizWand

Mathematical Reasoning on MATH500 (Accuracy Only)

92.1Accuracy (MATH500)GRPO
21.629639.924858.2276.5152Jun 30, 2026Jul 1, 2026Jul 2, 2026Jul 3, 2026Jul 4, 2026Jul 5, 2026Jul 6, 2026
Updated 3mo ago

Evaluation results

MethodLinks
GRPOTeacher model=Qwen3-30B-A3B-Instruct-25072026.0792.1
OPDTeacher model=Qwen3-30B-A3B-Instruct-2507, Student model=Qwen3-8B2026.0789
ReOPDTeacher model=Qwen3-4B-Instruct-2507, Student model=Qwen3-4B-Instruct-25072026.0788.9
ReOPDTeacher model=Qwen3-30B-A3B-Instruct-2507, Student model=Qwen3-8B2026.0788.4
OPDTeacher model=Qwen3-4B-Instruct-2507, Student model=Qwen3-4B-Instruct-25072026.0786.9
ReOPDTeacher model=Qwen3-30B-A3B-Instruct-2507, Student model=Qwen3-4B-Instruct-25072026.0786.5
OPDTeacher model=Qwen3-8B, Student model=Qwen3-4B-Instruct-25072026.0785.8
GRPOTeacher model=Qwen3-4B-Instruct-25072026.0785.1
ReOPDTeacher model=Qwen3-8B, Student model=Qwen3-4B-Instruct-25072026.0784.7
GRPOTeacher model=Qwen3-8B2026.0784.3
OPDTeacher model=Qwen3-30B-A3B-Instruct-2507, Student model=Qwen3-4B-Instruct-25072026.0783.8
Cold StartTeacher model=Qwen3-30B-A3B-Instruct-2507, Student model=Qwen3-8B2026.0782
SFTTeacher model=Qwen3-4B-Instruct-2507, Student model=Qwen3-4B-Instruct-25072026.0780.9
SFTTeacher model=Qwen3-8B, Student model=Qwen3-4B-Instruct-25072026.0780.8
Cold StartTeacher model=Qwen3-4B-Instruct-2507, Student model=Qwen3-4B-Instruct-25072026.0779.1
BaseTeacher model=Qwen3-30B-A3B-Instruct-2507, Student model=Qwen3-8B2026.0764.8
BaseTeacher model=Qwen3-4B-Instruct-2507, Student model=Qwen3-4B-Instruct-25072026.0758.2
LoRA-RLMOBackbone=Qwen2.5-7B-Instruct, Fine-tuning algorithm=GRPO, Training dataset=DAPO-Math-17k, Rank=r=32, Evaluation protocol=pass@4, B0=0, A0=V ⊤ −r2026.0656.73
LoRA-RLPOBackbone=Qwen2.5-7B-Instruct, Fine-tuning algorithm=GRPO, Training dataset=DAPO-Math-17k, Rank=r=32, Evaluation protocol=pass@4, B0=0, A0=V ⊤ r2026.0655.6
MiLoRABackbone=Qwen2.5-7B-Instruct, Fine-tuning algorithm=GRPO, Training dataset=DAPO-Math-17k, Rank=r=32, Evaluation protocol=pass@4, B0=U−rΣ 1/2 −r, A0=Σ1/2 −r V ⊤ −r2026.0654.13
LoRABackbone=Qwen2.5-7B-Instruct, Fine-tuning algorithm=GRPO, Training dataset=DAPO-Math-17k, Rank=r=32, Evaluation protocol=pass@4, B0=0, A0=N (0, 1/n)2026.0652.8
PiSSABackbone=Qwen2.5-7B-Instruct, Fine-tuning algorithm=GRPO, Training dataset=DAPO-Math-17k, Rank=r=32, Evaluation protocol=pass@4, B0=UrΣ 1/2 r, A0=Σ1/2 r V ⊤ r2026.0650.33
dOPSDBase Model=Dream-7B-Instruct2026.0742.2
BaseBase Model=Dream-7B-Instruct2026.0738.97
GRPOBase Model=Dream-7B-Instruct2026.0737.86
OPSD (answer-only)Base Model=Dream-7B-Instruct2026.0737.23
SFTBase Model=Dream-7B-Instruct2026.0736.49
dOPSDBase Model=LLaDA-8B-Instruct2026.0736
OPSD (full-solution)Base Model=Dream-7B-Instruct2026.0735.72
GRPOBase Model=LLaDA-8B-Instruct2026.0732.77
OPSD (answer-only)Base Model=LLaDA-8B-Instruct2026.0731.45
BaseBase Model=LLaDA-8B-Instruct2026.0731.24
SFTBase Model=LLaDA-8B-Instruct2026.0730.12
OPSD (full-solution)Base Model=LLaDA-8B-Instruct2026.0724.34