List your company
ResearchFormerly WizWand

Mathematical Reasoning on MATH500 1.0 (test)

96.9AccuracyA3PO
20.428840.281960.13579.9881Sep 26, 2025Nov 3, 2025Dec 11, 2025Jan 18, 2026Feb 25, 2026Apr 4, 2026May 12, 2026
Updated 5mo ago

Evaluation results

MethodLinks
A3POModel=Deepseek-R1-Distill-Qwen-7B2025.1296.9
Lp-RegModel=Deepseek-R1-Distill-Qwen-7B2025.1296.2
DAPO w/ Fork TokensModel=Deepseek-R1-Distill-Qwen-7B2025.1295.8
W-REINFORCEModel=Deepseek-R1-Distill-Qwen-7B2025.1295.7
DAPOModel=Deepseek-R1-Distill-Qwen-7B2025.1295.5
GRPOModel=Deepseek-R1-Distill-Qwen-7B2025.1295.2
A3POModel=Qwen3-8B-Base2025.1291.3
A3POModel=Qwen2.5-7B-Math2025.1290.4
Lp-RegModel=Qwen3-8B-Base2025.1287.4
W-REINFORCEModel=Qwen2.5-7B-Math2025.1287.3
Lp-RegModel=Qwen2.5-7B-Math2025.1287.1
NEMOTRON-NANO-12B-V2 (RLP + Post)Pre-training objective=RLP, Post-training pipeline=SFT + RLVR, Evaluation protocol=Pass@1 (average of 8 runs)2025.0987.05
W-REINFORCEModel=Qwen3-8B-Base2025.1286.9
DAPO w/ Fork TokensModel=Qwen2.5-7B-Math2025.1286.8
DAPO w/ Fork TokensModel=Qwen3-8B-Base2025.1286.2
DAPOModel=Qwen2.5-7B-Math2025.1285.4
DAPOModel=Qwen3-8B-Base2025.1284.5
GRPOModel=Qwen2.5-7B-Math2025.1283.8
NEMOTRON-NANO-12B-V2 (Base + Post)Pre-training objective=Next-token prediction, Post-training pipeline=SFT + RLVR, Evaluation protocol=Pass@1 (average of 8 runs)2025.0983.47
GRPOModel=Qwen3-8B-Base2025.1282.3
NEMOTRON-NANO-12B-V2 (Base)Pre-training objective=Next-token prediction, Post-training pipeline=None, Evaluation protocol=Pass@1 (average of 8 runs)2025.0979.95
NEMOTRON-NANO-12B-V2 (RLP)Pre-training objective=RLP, Post-training pipeline=None, Evaluation protocol=Pass@1 (average of 8 runs)2025.0978.68
polyGRPOBackbone=Qwen2.5-7B-Instruct2026.0464.94
GRPOBackbone=Qwen2.5-7B-Instruct2026.0463.2
LIDRBackbone=Qwen2.5-7B-Instruct2026.0462.46
MAPOBackbone=Qwen2.5-7B-Instruct2026.0461.2
BaseBackbone=Qwen2.5-7B-Instruct2026.0457.89
xRFTBackbone=Qwen2.5-7B-Instruct2026.0453.12
Llama3.1-8B-Instruct# data=N.A.2026.0251.9
SFT-Decomp + Curriculum# data=4500, backbone=Qwen2.5-1.5B2026.0251.6
SFT-Decomp# data=4500, backbone=Qwen2.5-1.5B2026.0250.8
SFT-MuggleMath# data=147787, backbone=Qwen2.5-1.5B2026.0250.4
SFT-Direct Distillation# data=360, backbone=Qwen2.5-1.5B2026.0249.6
SFT# data=360, backbone=Qwen2.5-1.5B2026.0248.4
SFT (full dataset)# data=7500, backbone=Qwen2.5-1.5B2026.0247.6
Base# data=N.A., backbone=Qwen2.5-1.5B2026.0247.2
MiniCPM3-4B# data=N.A.2026.0246.6
Gemma2-9B-Instruct# data=N.A.2026.0244.3
Gemma2-27B# data=N.A.2026.0242.7
Qwen2.5-3B# data=N.A.2026.0242.6
Llama-3-70B# data=N.A.2026.0242.5
TABOMSFT Training Domain=Code SFT (Ling-Coder-SFT)2026.0542.4
Mixtral-8x22B# data=N.A.2026.0241.7
SFT-SDSFT Training Domain=Code SFT (Ling-Coder-SFT)2026.0541.6
TABOMSFT Training Domain=Math SFT (MixChain-Z-PRM12K)2026.0541.1
No-SFTSFT Training Domain=None (Base Model)2026.0539.8
SFT-SDSFT Training Domain=Math SFT (MixChain-Z-PRM12K)2026.0539.8
SFT-GTSFT Training Domain=Math SFT (MixChain-Z-PRM12K)2026.0537.4
SFT-MetaMATH-Aug# data=2638, backbone=Qwen2.5-1.5B2026.0237.2
Qwen2.5-1.5B# data=N.A.2026.0235
SFT-GTSFT Training Domain=Code SFT (Ling-Coder-SFT)2026.0532.4
polyGRPOBackbone=Llama3-8B-Instruct2026.0426.71
BaseBackbone=Llama3-8B-Instruct2026.0426
MAPOBackbone=Llama3-8B-Instruct2026.0425.26
GRPOBackbone=Llama3-8B-Instruct2026.0424.43
LIDRBackbone=Llama3-8B-Instruct2026.0424.31
xRFTBackbone=Llama3-8B-Instruct2026.0423.37