List your company
ResearchFormerly WizWand

Mathematical Reasoning on MAWPS

98.5AccuracyOri-SFT
93.09294.49695.997.304Jan 25, 2024Jun 4, 2024Oct 13, 2024Feb 22, 2025Jul 3, 2025Nov 11, 2025Mar 23, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Ori-SFTModel=Qwen2.5-Math-7B-Instruct, GSM Train Method=Ori-SFT, Evaluation Protocol=Zero-shot2025.0698.5——
AbstRaLModel=Qwen2.5-Math-7B-Instruct, GSM Train Method=AbstRaL, Evaluation Protocol=Zero-shot2025.0698.5——
CoAModel=Qwen2.5-Math-7B-Instruct, GSM Train Method=CoA, Evaluation Protocol=Zero-shot2025.0698.3——
CoT-RLModel=Qwen2.5-Math-7B-Instruct, GSM Train Method=CoT-RL, Evaluation Protocol=Zero-shot2025.0698.1——
COMT+CCRLBackbone=QWEN3-4B2026.0197.9——
COMTModel=QWEN3-4B, Method=COMT2026.0197.9——
COMTBackbone=QWEN3-8B2026.0197.8——
GPT-4Protocol=PAL, Setting=Tool manipulation2024.0597.7——
GPT-4 (Code Interpreter)Decoding Strategy=Greedy2024.0297.7——
FEW-SHOTBackbone=QWEN2.5-7B2026.0197.6——
QWEN2.5-7BModel=QWEN2.5-7B, Method=FEW-SHOT2026.0197.6——
COMT+CCRLModel=QWEN3-4B, Method=COMT+CCRL2026.0197.6——
Base LLMBackbone=Qwen2.5-7B, Mode=Few-shot2026.0197.6——
Base LLMBackbone=Qwen2.5-7B, Mode=Few-shot2026.0197.6——
L12026.0397.596910.06
ROM2026.0397.572313.49
ROM_CSCCounterfactual Self-Correction=true2026.0397.552718.5
JiuZhang3.0-8BSetting=Tool manipulation, Model Size=8B2024.0597.3——
JiuZhang3.0-7BModel Scale=7B, Backbone=Mistral-7B2024.0597.3——
COMT+CCRLModel=QWEN2.5-7B, Method=COMT+CCRL2026.0197.3——
MAmmoTH2-8B-PlusModel Scale=8B2024.0597.1——
JiuZhang3.0-8BModel Scale=8B, Backbone=LLaMA-3-8B2024.0597.1——
MAmmoTH2-8x7B-PlusModel Scale=8x7B2024.0597——
FEW-SHOTBackbone=LLAMA3.1-8B2026.0197——
COMT+CCRLBackbone=QWEN2.5-7B2026.0197——
LLAMA3.1-8BModel=LLAMA3.1-8B, Method=FEW-SHOT2026.0197——
Base LLMBackbone=LLaMA3.1-8B, Mode=Few-shot2026.0197——
Base LLMBackbone=LLaMA3.1-8B, Mode=Few-shot2026.0197——
ChatGPT2024.0596.9——
COMTModel=QWEN2.5-7B, Method=COMT2026.0196.8——
FSLRBackbone=Qwen2-1.5B, Teacher Source=Self2026.0196.8——
FSLRBackbone=Qwen3-4B, Data Source=Self, Training Set=SVAMP2026.0196.8——
JiuZhang3.0-8x7BModel Scale=8x7B2024.0596.7——
COMTBackbone=QWEN3-4B2026.0196.7——
FSLRBackbone=Qwen2-1.5B, Teacher Source=Qwen2026.0196.7——
EAT2026.0396.71,1978.08
Vanilla2026.0396.671,2677.63
Cut2048max tokens=20482026.0396.671,2657.64
Cut1024max tokens=10242026.0396.671,2657.64
JiuZhang3.0-7BSetting=Tool manipulation, Model Size=7B2024.0596.6——
GPT-42024.0596.6——
FEW-SHOTBackbone=QWEN3-8B2026.0196.6——
QWEN3-8BModel=QWEN3-8B, Method=FEW-SHOT2026.0196.6——
CoT-SFTBackbone=Qwen2-1.5B, Teacher Source=Qwen2026.0196.6——
COMT+CCRLBackbone=QWEN3-8B2026.0196.5——
FSLRBackbone=Qwen3-4B, Data Source=Qwen, Training Set=SVAMP2026.0196.2——
CoT-SFTBackbone=Qwen3-4B, Data Source=Self, Training Set=SVAMP2026.0196.2——
FSLRBackbone=Qwen3-4B, Data Source=LLaMA, Training Set=SVAMP2026.0196——
FEW-SHOTBackbone=QWEN3-4B2026.0195.9——
QWEN3-4BModel=QWEN3-4B, Method=FEW-SHOT2026.0195.9——
Base LLMBackbone=Qwen2-1.5B, Mode=Few-shot2026.0195.9——
Base LLMBackbone=Qwen3-4B, Mode=Few-shot2026.0195.9——
Cut512max tokens=5122026.0395.831,2657.58
Qwen-1.5-72BModel Scale=72B2024.0595.8——
MAmmoTH2-7B-PlusModel Scale=7B2024.0595.8——
COMTModel=QWEN3-8B, Method=COMT2026.0195.8——
DeepSeekMath-7B-InstructModel Scale=7B2024.0595.7——
OpenMath-CodeLlamaModel Size=70B, Base Model=CodeLlama, Decoding Strategy=Greedy2024.0295.7——
OpenMath-Llama2Model Size=70B, Base Model=Llama-2, Decoding Strategy=Greedy2024.0295.6——
DeepSeekMath-7B-RLModel Scale=7B2024.0595.5——
COMT+CCRLModel=LLAMA3.1-8B, Method=COMT+CCRL2026.0195.5——
OpenMath-Mistral-7BSetting=Tool manipulation, Model Size=7B, Backbone=Mistral2024.0595.4——
OpenMath-Mistral-7BModel Size=7B, Base Model=Mistral, Decoding Strategy=Greedy2024.0295.4——
COMT+CCRLBackbone=LLAMA3.1-8B2026.0195.4——
COMTBackbone=QWEN2.5-7B2026.0195.3——
GPTAugBackbone=Qwen2-Math-7B, # Samples=88.62k2025.0395.2——
COT-SFT+RLModel=QWEN2.5-7B, Method=COT-SFT+RL2026.0195.1——
COMT+CCRLModel=QWEN3-8B, Method=COMT+CCRL2026.0195——
FSLRBackbone=LLaMA3.1-8B, Teacher Source=Self2026.0195——
OpenMath-CodeLlamaModel Size=34B, Base Model=CodeLlama, Decoding Strategy=Greedy2024.0294.9——
KPMath-DSMath-7BModel Scale=7B2024.0594.8——
LEMMABackbone=Qwen2-Math-7B, # Samples=88.90k2025.0394.8——
FSLRBackbone=Qwen2.5-7B, Data Source=LLaMA, Training Set=SVAMP2026.0194.8——
FSLRBackbone=Qwen2.5-7B, Teacher Source=Qwen2026.0194.6——
Rho-1-Math-7B-CodeSetting=Tool manipulation, Model Size=7B2024.0594.5——
Intern-Math-20BModel Scale=20B2024.0594.4——
FSLRBackbone=LLaMA3.1-8B, Data Source=LLaMA, Training Set=SVAMP2026.0194.4——
Qwen-1.5-110BModel Scale=110B2024.0594.3——
CoT-SFTBackbone=Qwen3-4B, Data Source=Qwen, Training Set=SVAMP2026.0194.3——
FSLRBackbone=LLaMA3.1-8B, Data Source=Qwen, Training Set=SVAMP2026.0194.2——
FSLRBackbone=LLaMA3.1-8B, Data Source=Self, Training Set=SVAMP2026.0194.2——
Llama-3-SynEevaluation_mode=Few-shot2024.0794.1——
FSLRBackbone=Qwen2.5-7B, Teacher Source=LLaMA2026.0194.1——
Rho-1-Math-7BModel Scale=7B2024.0594——
COMTBackbone=LLAMA3.1-8B2026.0194——
FSLRBackbone=Qwen2.5-7B, Data Source=Qwen, Training Set=SVAMP2026.0194——
Mixtral-8x7BModel Scale=8x7B2024.0593.9——
WizardMath-7B-1.1Model Scale=7B2024.0593.8——
ToRAModel Size=70B, Base Model=Llama-2, Decoding Strategy=Greedy2024.0293.8——
COT-SFT+RLModel=QWEN3-4B, Method=COT-SFT+RL2026.0193.8——
FSLRBackbone=LLaMA3.1-8B, Teacher Source=Qwen2026.0193.8——
MetaMath-Mistral-7BModel Scale=7B2024.0593.7——
COT-SFT+RLBackbone=QWEN3-4B2026.0193.7——
COT-SFTModel=QWEN3-4B, Method=COT-SFT2026.0193.7——
OpenMath-CodeLlamaModel Size=13B, Base Model=CodeLlama, Decoding Strategy=Greedy2024.0293.6——
Abel-7B-002Model Scale=7B2024.0593.5——
OpenMath-CodeLlamaModel Size=7B, Base Model=CodeLlama, Decoding Strategy=Greedy2024.0293.5——
COMTModel=LLAMA3.1-8B, Method=COMT2026.0193.5——
DeepSeek-Coder-BaseSize=33B, Protocol=Program-Aided Math Reasoning (PAL)2024.0193.3——
ToRAModel Size=34B, Base Model=CodeLlama, Decoding Strategy=Greedy2024.0293.3——