List your company
ResearchFormerly WizWand

Mathematical Reasoning on MATH (Accuracy and Token Usage)

96.67AccuracyFull-Graph
63.608472.191780.77589.3583Mar 7, 2024Jul 7, 2024Nov 6, 2024Mar 8, 2025Jul 8, 2025Nov 7, 2025Mar 9, 2026
Updated 7mo ago

Evaluation results

MethodLinks
Full-GraphMulti=true, Topo=false, Role=false, Node=false, Backbone=Qwen3-80B2026.0296.67——
HieraMASMulti=true, Topo=true, Role=true, Node=true, Backbone=GPT-5-Mini2026.0296.67——
Full-GraphMulti=true, Topo=false, Role=false, Node=false, Backbone=GPT-5-Mini2026.0295.56——
AFlowMulti=true, Topo=true, Role=false, Node=false, Backbone=GPT-5-Mini2026.0295.56——
HieraMASMulti=true, Topo=true, Role=true, Node=true, Backbone=Qwen3-80B2026.0295.56——
Self-Consistency+CoTMulti=true, Topo=false, Role=false, Node=false, Backbone=GPT-5-Mini2026.0294.44——
LLM-DebateMulti=true, Topo=false, Role=false, Node=false, Backbone=GPT-5-Mini2026.0294.44——
LLM-DebateMulti=true, Topo=false, Role=false, Node=false, Backbone=Qwen3-80B2026.0294.44——
Random-GraphMulti=true, Topo=false, Role=false, Node=false, Backbone=Qwen3-80B2026.0294.44——
4B-InstructModel=4B-Instruct, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0394.3——
4B-InstructModel=4B-Instruct, RL Algorithm=GSPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0394——
4B-InstructModel=4B-Instruct, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0393.9——
4B-InstructModel=4B-Instruct, RL Algorithm=Base, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0393.7——
Self-ConsistencyMulti=true, Topo=false, Role=false, Node=false, Backbone=GPT-5-Mini2026.0293.33——
Random-GraphMulti=true, Topo=false, Role=false, Node=false, Backbone=GPT-5-Mini2026.0293.33——
4B-InstructModel=4B-Instruct, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0393.3——
CoTMulti=false, Topo=false, Role=false, Node=false, Backbone=GPT-5-Mini2026.0292.22——
Self-Consistency+CoTMulti=true, Topo=false, Role=false, Node=false, Backbone=Qwen3-80B2026.0292.22——
Self-ConsistencyMulti=true, Topo=false, Role=false, Node=false, Backbone=Qwen3-80B2026.0291.11——
GDesignerMulti=true, Topo=true, Role=false, Node=false, Backbone=GPT-5-Mini2026.0291.11——
MASRouterMulti=true, Topo=true, Role=true, Node=false, Backbone=GPT-5-Mini2026.0291.11——
CODABackbone=Qwen3-14B, Sampling Strategy=mean@82026.0391.11,605—
4BModel=4B, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0390.5——
VLPBackbone=Qwen3-14B, Sampling Strategy=mean@82026.0390.41,821—
GRPOBackbone=Qwen3-14B, Sampling Strategy=mean@82026.0390.32,229—
CoTMulti=false, Topo=false, Role=false, Node=false, Backbone=Qwen3-80B2026.0290——
ASRRBackbone=Qwen3-14B, Sampling Strategy=mean@82026.0389.11,324—
4BModel=4B, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0388.9——
MASRouterMulti=true, Topo=true, Role=true, Node=false, Backbone=Qwen3-80B2026.0288.88——
GRPOBackbone=Qwen3-8B, Sampling Strategy=mean@82026.0387.82,270—
GDesignerMulti=true, Topo=true, Role=false, Node=false, Backbone=Qwen3-80B2026.0287.77——
CODABackbone=Qwen3-8B, Sampling Strategy=mean@82026.0387.61,700—
4BModel=4B, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0387.5——
VLPBackbone=Qwen3-8B, Sampling Strategy=mean@82026.0387.51,801—
4BModel=4B, RL Algorithm=GSPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0387——
ASRRBackbone=Qwen3-8B, Sampling Strategy=mean@82026.0386.71,598—
GRPOBackbone=Qwen3-4B, Sampling Strategy=mean@82026.0385.32,774—
CODABackbone=Qwen3-4B, Sampling Strategy=mean@82026.0385.11,914—
ASRRBackbone=Qwen3-4B, Sampling Strategy=mean@82026.0384.81,464—
4B-InstructModel=4B-Instruct, RL Algorithm=Naive, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0384.5——
VLPBackbone=Qwen3-4B, Sampling Strategy=mean@82026.0384.51,716—
AFlowMulti=true, Topo=true, Role=false, Node=false, Backbone=Qwen3-80B2026.0284.44——
AbstRaLModel=Qwen2.5-Math-7B-Instruct, GSM Train Method=AbstRaL, Evaluation Protocol=Zero-shot2025.0683.9——
CoAModel=Qwen2.5-Math-7B-Instruct, GSM Train Method=CoA, Evaluation Protocol=Zero-shot2025.0683.7——
4BModel=4B, RL Algorithm=Base, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0383.6——
Ori-SFTModel=Qwen2.5-Math-7B-Instruct, GSM Train Method=Ori-SFT, Evaluation Protocol=Zero-shot2025.0683.5——
CoT-RLModel=Qwen2.5-Math-7B-Instruct, GSM Train Method=CoT-RL, Evaluation Protocol=Zero-shot2025.0682.9——
HACPOModel Backbone=Qwen3-8B-Base2026.0381.3——
ROSA2Model=Qwen3-8B2026.0380.8——
Zero-shot-EIModel=Gemini-1.5-Pro, Model Role=Teacher Model2026.0280.59——
HACPOModel Backbone=Qwen3-4B-Base2026.0380.3——
4B-BaseModel=4B-Base, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0380.1——
4B-BaseModel=4B-Base, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0379.4——
qwen3-4BModel=qwen3-4B, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0379.4——
GRPOModel Backbone=Qwen3-4B-Base2026.0378.8——
4B-BaseModel=4B-Base, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0378.8——
qwen3-4BModel=qwen3-4B, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0378.8——
4B-BaseModel=4B-Base, RL Algorithm=GSPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0378.7——
qwen3-4BModel=qwen3-4B, RL Algorithm=GSPO, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0378.7——
qwen3-4BModel=qwen3-4B, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0378.3——
BaseMulti=false, Topo=false, Role=false, Node=false, Backbone=GPT-5-Mini2026.0277.78——
Zero-shotModel=Gemini-1.5-Pro, Model Role=Teacher Model2026.0277.2——
ExGRPOModel=Qwen2.5-7B-Instruct, Model Role=Student Model, Adapter=On-Policy Adapt.2026.0277.12——
ExGRPOModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.0276.58——
RevThinkModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.0276.4——
On-Policy DistillationModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.0276.24——
Answer AugModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.0276.18——
Divide-or-ConquerModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.0276.05——
Question AugModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.0275.97——
Rephrase QuestionModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.0275.64——
Distill Step-by-StepModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.0275.32——
SKDModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.0274.88——
BaseMulti=false, Topo=false, Role=false, Node=false, Backbone=Qwen3-80B2026.0274.44——
Zero-shotModel=Qwen2.5-7B-Instruct, Model Role=Student Model2026.0274.4——
Llama 3 405Bshot=0-shot, CoT=true2024.0773.8——
4BModel=4B, RL Algorithm=Naive, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0373.7——
M2CLBackbone=Qwen-72B, Number of LLMs=82026.0272.5——
Qwen3-14B-BaseBackbone=Qwen3-14B, Sampling Strategy=mean@82026.0372.5914—
4B-BaseModel=4B-Base, RL Algorithm=Naive, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0371.2——
qwen3-4BModel=qwen3-4B, RL Algorithm=Naive, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0371.2——
Claude 3.5 Sonnetshot=0-shot, CoT=false2024.0771.1——
ROSA2Model=Qwen3-0.6B-Instruct2026.0370.8——
ROSA2Model=Qwen2.5-7B-Base2026.0368.4——
BaseModel Backbone=Qwen3-4B-Base2026.0367.6——
4B-BaseModel=4B-Base, RL Algorithm=Base, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0367.6——
qwen3-4BModel=qwen3-4B, RL Algorithm=Base, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0367.6——
ROSAModel=Qwen3-8B, Update Location=LM Head, Reward Model=Model-based2025.0967.4——
1.7B-BaseModel=1.7B-Base, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0367.4——
Qwen3-8B-BaseBackbone=Qwen3-8B, Sampling Strategy=mean@82026.0367.3851—
ROSAModel=Qwen3-0.6B-Instruct2026.0366.2——
Llama 3.1 InstructModel Scale=70B2025.0466.2——
GPT-42024.0366.1——
HACPOModel Backbone=Qwen3-1.7B-Base2026.0366.1——
ROSAModel=Qwen3-8B, Update Location=LM Head, Reward Model=Rule-based2025.0965.8——
ROSAModel=Qwen3-8B, Update Location=Hidden States, Reward Model=Rule-based2025.0965.8——
ROSAModel=Qwen2.5-7B-Instruct, Update Location=LM Head, Reward Model=Model-based2025.0965.2——
1.7B-BaseModel=1.7B-Base, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0365.2——
TextGradModel=Qwen3-0.6B-Instruct2026.0365——
1.7B-BaseModel=1.7B-Base, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0365——
ℓ-MoEaccActive Params=8.44B, Total Params=94.8B2026.0164.88——