List your company
ResearchFormerly WizWand

Mathematical Reasoning on MATHEMATICS (Accuracy)

74.1AccuracyLEMMA
5.66823.43441.258.966Mar 12, 2024Jul 17, 2024Nov 21, 2024Mar 28, 2025Aug 2, 2025Dec 7, 2025Apr 13, 2026
Updated 6mo ago

Evaluation results

MethodLinks
LEMMABackbone=Qwen2-Math-7B, # Samples=88.90k2025.0374.1
RefAug-90kBackbone=Qwen2-Math-7B, # Samples=89.92k2025.0368.7
GPTAugBackbone=Qwen2-Math-7B, # Samples=88.62k2025.0364.9
RefAugBackbone=Qwen2-Math-7B, # Samples=29.94k2025.0362.7
S2LTarget Model=Phi-3-Mini (3.8B), Fine-tuning Data=50%, Evaluation Protocol=Zero-shot2024.0362.1
MetaMathBackbone=Qwen2-Math-7B, # Samples=394.99k2025.0360.7
Full TrainingTarget Model=Phi-3-Mini (3.8B), Fine-tuning Data=100%, Evaluation Protocol=Zero-shot2024.0360.2
RFTBackbone=Qwen2-Math-7B, # Samples=86.52k2025.0357.4
ISCBackbone=Qwen2-Math-7B, # Samples=86.78k2025.0351.9
SFTBackbone=Qwen2-Math-7B, # Samples=14.97k2025.0350.3
S2LTarget Model=Llama-2-7B, Fine-tuning Data=50%, Evaluation Protocol=Zero-shot2024.0345.2
PretrainedTarget Model=Phi-3-Mini (3.8B), Evaluation Protocol=Zero-shot2024.0344.3
Full TrainingTarget Model=Llama-2-7B, Fine-tuning Data=262K, Evaluation Protocol=Zero-shot2024.0343.9
LEMMABackbone=Mistral-7B-v0.1, # Samples=88.90k2025.0340.3
GREATSSelection Protocol=batch-wise2026.0438.19
GradNormSelection Protocol=source-wise2026.0438
COLMSelection Protocol=source-wise2026.0438
COLMSelection Protocol=batch-wise2026.0437.4
UniPROTSelection Protocol=batch-wise2026.0437.2
UniPROTSelection Protocol=source-wise2026.0436.9
GradNormSelection Protocol=batch-wise2026.0436.5
GREATSSelection Protocol=source-wise2026.0435.6
RefAug-90kBackbone=Mistral-7B-v0.1, # Samples=89.92k2025.0334.8
GPTAugBackbone=Mistral-7B-v0.1, # Samples=88.62k2025.0334.6
SBERTSelection Protocol=source-wise2026.0434.05
SBERTSelection Protocol=batch-wise2026.0434
RefAugBackbone=Mistral-7B-v0.1, # Samples=29.94k2025.0333.7
FTBase Model=PHI-32026.0433.3
MaxLossSelection Protocol=source-wise2026.0431.45
MaxLossSelection Protocol=batch-wise2026.0430.06
ISCBackbone=Mistral-7B-v0.1, # Samples=86.78k2025.0327.4
MetaMathBackbone=Mistral-7B-v0.1, # Samples=394.99k2025.0326.6
GREATSSelection Strategy=source-wise2026.0421.1
UniPROTSelection Strategy=source-wise2026.0420.9
GREATSSelection Strategy=batch-wise2026.0420.7
GradNormSelection Strategy=batch-wise2026.0420.4
GradNormSelection Strategy=source-wise2026.0420.2
COLMSelection Strategy=batch-wise2026.0419.8
UniPROTSelection Strategy=batch-wise2026.0419.8
FT2026.0419.4
COLMSelection Strategy=source-wise2026.0418.9
SBERTSelection Strategy=source-wise2026.0418.1
SFTBackbone=Mistral-7B-v0.1, # Samples=14.97k2025.0316.6
RFTBackbone=Mistral-7B-v0.1, # Samples=86.52k2025.0316.6
SBERTSelection Strategy=batch-wise2026.0416.6
PretrainedTarget Model=Llama-2-7B, Evaluation Protocol=Zero-shot2024.038.3