List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME25 (Pass@1)

72.5Pass@1RePO
-2.916.67536.2555.825Feb 1, 2026Feb 17, 2026Mar 6, 2026Mar 23, 2026Apr 8, 2026Apr 25, 2026May 12, 2026
Updated 5mo ago

Evaluation results

MethodLinks
RePOtraining_data=SuperGPQA subset (11k)2026.0272.5
Qwen3-8Btraining_data=SuperGPQA subset (11k)2026.0266.4
GRPOtraining_data=SuperGPQA subset (11k)2026.0265.8
LUFFYtraining_data=SuperGPQA subset (11k)2026.0264.1
Qwen3-4B + SCRLMax Generation Length=15,3602026.0360.2
Qwen3-4B + TTRLMax Generation Length=15,3602026.0359.3
Qwen3-4B + SCRLMax Generation Length=10,2402026.0353.6
Qwen3-4BMax Generation Length=15,3602026.0351.9
Qwen3-4B + TTRLMax Generation Length=10,2402026.0350.9
Qwen3-1.7BModel Name=Qwen3-1.7B, Model Size=1.7B, Method Type=Baseline2026.0447.5
Baseline ModelRollout Length=16K tokens2026.0547.1
PieceHint-Nemotron-1.5BModel Name=Nemotron-1.5B, Model Size=1.5B, Method Type=PieceHint2026.0443.7
DeepSeek-R1-Distill-32BModel Name=DeepSeek-R1-Distill-32B, Model Size=32B, Method Type=Baseline2026.0439.5
Qwen3-4BMax Generation Length=10,2402026.0338.9
PieceHint-Qwen3-1.7BModel Name=Qwen3-1.7B, Model Size=1.7B, Method Type=PieceHint2026.0438.8
Qwen3-4BModel Name=Qwen3-4B, Model Size=4B, Method Type=Baseline2026.0437.3
Nemotron-1.5BModel Name=Nemotron-1.5B, Model Size=1.5B, Method Type=Baseline2026.0434.2
DeepSeek-R1-Distill-7BModel Name=DeepSeek-R1-Distill-7B, Model Size=7B, Method Type=Baseline2026.0429.9
DAPO + PIPOModel=Qwen3-4B-Base2026.0429.6
GRPO + PIPOModel=Qwen3-8B-Base2026.0429.6
GSPO + PIPOModel=Qwen3-8B-Base2026.0429.6
DAPO + PIPOModel=Qwen3-8B-Base2026.0429.6
GSPO + PIPOModel=Qwen3-4B-Base2026.0425.9
GRPOModel=Qwen3-8B-Base2026.0425.9
DAPOModel=Qwen3-8B-Base2026.0425.9
PieceHint-DeepSeek-1.5BModel Name=DeepSeek-R1-Distill-1.5B, Model Size=1.5B, Method Type=PieceHint2026.0425.5
FP8 Rollout P3ORollout Length=16K tokens, Training Iteration=30, Rollout Precision=FP8, Training Precision=BF162026.0525.4
DAPOModel=Qwen3-4B-Base2026.0425
FP8 Rollout GRPORollout Length=16K tokens, Training Iteration=15, Rollout Precision=FP8, Training Precision=BF162026.0525
FP8 Rollout P3ORollout Length=16K tokens, Training Iteration=15, Rollout Precision=FP8, Training Precision=BF162026.0523.7
GRPO + PIPOModel=Qwen3-4B-Base2026.0422.2
GSPOModel=Qwen3-8B-Base2026.0422.2
DeepSeek-R1-Distill-1.5BModel Name=DeepSeek-R1-Distill-1.5B, Model Size=1.5B, Method Type=Baseline2026.0421.8
GRPOModel=Qwen3-4B-Base2026.0418.5
GSPOModel=Qwen3-4B-Base2026.0418.5
P3ORollout Length=4K tokens, Clipping Parameter (epsilon)=∈ {0.2, 0.4, 0.6}2026.0518.3
GRPO (clip avg)Rollout Length=4K tokens, Clipping Parameter (epsilon)=∈ {0.2, 0.4, 0.6}2026.0516
SetPO+GRPOBase Model=Qwen2.5-Math-7B, Optimization Algorithm=GRPO2026.0213.6
SetPO+DAPOBase Model=Qwen2.5-Math-7B, Optimization Algorithm=DAPO2026.0213.5
Base ModelModel=Qwen3-8B-Base2026.0411.1
DAPOBase Model=Qwen2.5-Math-7B2026.0211
SetPO+GSPOBase Model=Qwen2.5-Math-7B, Optimization Algorithm=GSPO2026.0210.1
GRPOBase Model=Qwen2.5-Math-7B2026.029.7
GSPOBase Model=Qwen2.5-Math-7B2026.027.9
Base ModelModel=Qwen3-4B-Base2026.047.4
Qwen2.5-Math-7BModel Type=Base Model2026.024
Baseline ModelRollout Length=4K tokens2026.053.3
FP8 Rollout GRPORollout Length=16K tokens, Training Iteration=30, Rollout Precision=FP8, Training Precision=BF162026.050