List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME 2025 (Accuracy, Avg, Speedup)

99.2AccuracyTRICE-30B
36.38452.6926985.308Oct 8, 2025Nov 20, 2025Jan 2, 2026Feb 14, 2026Mar 29, 2026May 11, 2026Jun 23, 2026
Updated 4mo ago

Evaluation results

MethodLinks
TRICE-30BTool Usage=true, Parameter Scale=~30B, Evaluation Protocol=unified2026.0599.2——
GLM-4.7-Flash w/ recipeTool Usage=true, Parameter Scale=~30B, Evaluation Protocol=unified2026.0598.3——
TRICE-4BTool Usage=true, Parameter Scale=<10B, Evaluation Protocol=unified2026.0596.7——
Nemotron-3-Nano-30B-A3BTool Usage=true, Parameter Scale=~30B, Evaluation Protocol=unified2026.0596.7——
GLM-5Params=744B2026.0696.7——
VibeThinker-3B + CLRParams=3B, Post-training optimization=CLR2026.0696.7——
Kimi K2.5Params=1T2026.0696.1——
Gemini 3 ProParams=N/A2026.0696——
GLM-4.7-FlashTool Usage=true, Parameter Scale=~30B, Evaluation Protocol=unified2026.0595——
GPT-5 (high)Params=N/A2026.0694.6——
Qwen3.5-35B-A3BTool Usage=false, Parameter Scale=~30B, Evaluation Protocol=unified2026.0594.2——
MiMo v2 FlashParams=309B2026.0694.1——
Qwen3.6 PlusParams=N/A2026.0693.3——
DeepSeek V3.2Params=671B2026.0693.1——
Claude Opus 4.5Params=N/A2026.0692.8——
GPT-OSS (high)Params=120B2026.0692.5——
Grok 4Params=N/A2026.0691.7——
VibeThinker-3BParams=3B2026.0691.4——
ASTER-4B†Tool Usage=true, Parameter Scale=<10B, Evaluation Protocol=original2026.0590——
VeryTraceModel=Q3-80b-T2026.0690——
TRICE-30BTool Usage=false, Parameter Scale=~30B, Evaluation Protocol=unified2026.0589.2——
OpenAI o3 (high)Params=N/A2026.0688.9——
Qwen3-30B-A3B-Thinking-2507Tool Usage=false, Parameter Scale=~30B, Evaluation Protocol=unified2026.0588.8——
DeepSeek R1 0528Params=671B2026.0687.5——
GPT-OSS-20BTool Usage=true, Parameter Scale=~30B, Evaluation Protocol=unified2026.0586.7——
Gemini 2.5 ProParams=N/A2026.0686.7——
CoVeModel=Q3-80b-T2026.0686.67——
AgentMath-30B-A3B†Tool Usage=true, Parameter Scale=~30B, Evaluation Protocol=original2026.0586.4——
Qwen3.5-9BTool Usage=false, Parameter Scale=<10B, Evaluation Protocol=unified2026.0585.8——
AgentMath-8B†Tool Usage=true, Parameter Scale=<10B, Evaluation Protocol=original2026.0584.7——
VanillaModel=Q3-80b-T2026.0683.33——
NPModel=Q3-80b-T2026.0683.33——
Qwen3-4B-Thinking-2507Tool Usage=false, Parameter Scale=<10B, Evaluation Protocol=unified2026.0582.5——
VeryTraceModel=Q3-80b-I2026.0680——
TRICE-4BTool Usage=false, Parameter Scale=<10B, Evaluation Protocol=unified2026.0579.2——
Qwen3.5-4BTool Usage=false, Parameter Scale=<10B, Evaluation Protocol=unified2026.0575.8——
SeirênesBackbone=Qwen3-30B-A3B, same budget=true2026.0574.1——
CoVeModel=Q3-80b-I2026.0673.33——
NPModel=Q3-80b-I2026.0673.33——
Gemini 2.5 FlashParams=N/A2026.0672——
Glob. non-uniBackbone Model=Qwen3-8B2025.1070——
Loc. uniBackbone Model=Qwen3-8B2025.1069——
Qwen3-30B-A3B-Instruct-2507Tool Usage=false, Parameter Scale=~30B, Evaluation Protocol=unified2026.0567.5——
Qwen3-4B-LambdaGRPOModel Architecture=Qwen3, Model Size=4B, Training Algorithm=LambdaGRPO2026.0567.0876.49—
LamPOBackbone=Qwen3-4B2026.0567.08——
Mean Acc.Backbone Model=Qwen3-8B2025.1067——
DAPOBackbone=Qwen3-30B-A3B, same budget=true2026.0566.8——
ReLATModel=Qwen3-14B2026.0666.7——
Glob. uniBackbone Model=Qwen3-8B2025.1066——
Qwen3-4B-GRPOModel Architecture=Qwen3, Model Size=4B, Training Algorithm=GRPO2026.0565.8375.4—
GRPOBackbone=Qwen3-4B2026.0565.83——
SeirênesBackbone=Qwen3-4B-Instruct, same budget=true2026.0564.9——
Dr.GRPOBackbone=Qwen3-4B-Instruct, same budget=true2026.0563.3——
ReLATModel=Qwen3-8B2026.0663.3——
LatentMASModel=Qwen3-14B2026.0663.3——
Self-Cert.Backbone Model=Qwen3-8B2025.1063——
Loc. non-uniBackbone Model=Qwen3-8B2025.1063——
Qwen3-30B-A3BBackbone=Qwen3-30B-A3B2026.0560.6——
DAPOBackbone=Qwen3-4B-Instruct, same budget=true2026.0560.2——
High Conf.Backbone Model=Qwen3-8B2025.1060——
Low Ent.Backbone Model=Qwen3-8B2025.1060——
Qwen3-4BModel Architecture=Qwen3, Model Size=4B2026.056073.18—
CoTBackbone=Qwen3-4B2026.0560——
TextMASModel=Qwen3-14B2026.0660——
InTBackbone=Qwen3-4B-Instruct2026.0558.9——
ReLATModel=Qwen3-4B2026.0656.7——
SingleModel=Qwen3-14B2026.0656.7——
VanillaModel=Q3-80b-I2026.0653.33——
TextMASModel=Qwen3-8B2026.0653.3——
LatentMASModel=Qwen3-8B2026.0653.3——
Loc. uniBackbone Model=DS-R1-Distill-Qwen-7B2025.1053——
Glob. non-uniBackbone Model=DS-R1-Distill-Qwen-7B2025.1052——
LatentMASModel=Qwen3-4B2026.0650——
Scaf-GRPOBackbone=Qwen3-4B-Instruct2026.0549.1——
SAGEBackbone=Qwen3-4B-Instruct2026.0549.1——
Qwen3-4B-InstructBackbone=Qwen3-4B-Instruct2026.0548.8——
Self-Cert.Backbone Model=DS-R1-Distill-Qwen-7B2025.1048——
High Conf.Backbone Model=DS-R1-Distill-Qwen-7B2025.1048——
Low Ent.Backbone Model=DS-R1-Distill-Qwen-7B2025.1048——
SingleModel=Qwen3-8B2026.0646.7——
GRPO + ARROLBackbone=Qwen-3-4B-Base, Method=GRPO + ARROL2026.0346.6753.541.63
GRPO + ARROLBackbone=Qwen-3-8B-Base, Method=GRPO + ARROL2026.0346.6759.531.62
PieceHintBase model=DeepSeek-R1-Distill-7B2026.0443.9——
SCOPE-RLBase Model=Qwen3-4B2025.1043.7——
SingleModel=Qwen3-4B2026.0643.3——
TextMASModel=Qwen3-4B2026.0643.3——
Entropy-RegBase Model=Qwen3-4B, λ=0.012025.1042.6——
CISPOBase Model=Qwen3-4B2025.1042.6——
KL-covBase Model=Qwen3-4B2025.1042.3——
DAPOBase Model=Qwen3-4B2025.1041.7——
Entropy-AdvBase Model=Qwen3-4B2025.1041.7——
GRPOBase Model=Qwen3-4B2025.1041.5——
STARE-O1Scenario=Long CoT, Backbone Model=DeepSeek-R1-Distill-Qwen-1.5B, Model Scale=1.5B2026.0641.5——
STARE-C2Scenario=Long CoT, Backbone Model=DeepSeek-R1-Distill-Qwen-1.5B, Model Scale=1.5B2026.0640.5——
GRPOBackbone=Qwen-3-4B-Base, Method=GRPO2026.034051.01—
GRPOBackbone=Qwen-3-8B-Base, Method=GRPO2026.034056.66—
Mean Acc.Backbone Model=DS-R1-Distill-Qwen-7B2025.1040——
ReLATModel=DeepSeek-R1-Distill-Qwen-7B2026.0640——
Loc. uniBackbone Model=DS-R1-Distill-Llama-8B2025.1039——
JustRLScenario=Long CoT, Backbone Model=DeepSeek-R1-Distill-Qwen-1.5B, Model Scale=1.5B2026.0638.8——