List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME 25 (pass@1)

80pass@1CompassMax-V3-Thinking
-1.32819.78640.962.014Dec 8, 2025Dec 19, 2025Dec 30, 2025Jan 10, 2026Jan 21, 2026Feb 1, 2026Feb 12, 2026
Updated 8mo ago

Evaluation results

MethodLinks
CompassMax-V3-Thinking2025.1280
Chain of Mindset (CoM)Base Model=Qwen3-VL-32B-Instruct, Temperature=0.7, Top-p=0.95, Max Tokens=327682026.0273.33
DeepSeek-R12025.1270
ReActBase Model=Qwen3-VL-32B-Instruct, Temperature=0.7, Top-p=0.95, Max Tokens=327682026.0263.33
w/o MergingModels=Math2026.0261.25
Zero-shot CoTBase Model=Qwen3-VL-32B-Instruct, Temperature=0.7, Top-p=0.95, Max Tokens=327682026.0260
MRPBase Model=Qwen3-VL-32B-Instruct, Temperature=0.7, Top-p=0.95, Max Tokens=327682026.0260
w/o MergingModels=Code2026.0258.44
SCF-RKLModels=Fuse2026.0258.02
Direct I/OBase Model=Qwen3-VL-32B-Instruct, Temperature=0.7, Top-p=0.95, Max Tokens=327682026.0256.67
Chain of CodeBase Model=Qwen3-VL-32B-Instruct, Temperature=0.7, Top-p=0.95, Max Tokens=327682026.0256.67
SCEModels=Fuse2026.0256.25
Task ArithmeticModels=Fuse2026.0254.17
Dare Task ArithmeticModels=Fuse2026.0254.06
Ties MergingModels=Fuse2026.0252.81
Tree of ThoughtsBase Model=Qwen3-VL-32B-Instruct, Temperature=0.7, Top-p=0.95, Max Tokens=327682026.0250
Dare Ties MergingModels=Fuse2026.0249.27
Meta-ReasonerBase Model=Qwen3-VL-32B-Instruct, Temperature=0.7, Top-p=0.95, Max Tokens=327682026.0236.67
MinPROModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=1282026.0133.4
CoMBase Model=Gemini-2.0-Flash2026.0233.33
Chain of Mindset (CoM)Base Model=Gemini-2.0-Flash, Temperature=0.7, Top-p=0.95, Max Tokens=81922026.0233.33
CISPOModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=1282026.0132.7
M2POModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=1282026.0132
GSPOModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=1282026.0131.8
MinPROModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=1282026.0130.3
Chain of CodeBase Model=Gemini-2.0-Flash2026.0230
Chain of CodeBase Model=Gemini-2.0-Flash, Temperature=0.7, Top-p=0.95, Max Tokens=81922026.0230
Direct I/OBase Model=Gemini-2.0-Flash2026.0226.67
MRPBase Model=Gemini-2.0-Flash2026.0226.67
Meta-ReasonerBase Model=Gemini-2.0-Flash2026.0226.67
Direct I/OBase Model=Gemini-2.0-Flash, Temperature=0.7, Top-p=0.95, Max Tokens=81922026.0226.67
MRPBase Model=Gemini-2.0-Flash, Temperature=0.7, Top-p=0.95, Max Tokens=81922026.0226.67
Meta-ReasonerBase Model=Gemini-2.0-Flash, Temperature=0.7, Top-p=0.95, Max Tokens=81922026.0226.67
M2POModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=1282026.0125.8
CISPOModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=1282026.0125.1
Zero-shot CoTBase Model=Gemini-2.0-Flash2026.0223.33
Tree of ThoughtsBase Model=Gemini-2.0-Flash2026.0223.33
ReActBase Model=Gemini-2.0-Flash2026.0223.33
Zero-shot CoTBase Model=Gemini-2.0-Flash, Temperature=0.7, Top-p=0.95, Max Tokens=81922026.0223.33
Tree of ThoughtsBase Model=Gemini-2.0-Flash, Temperature=0.7, Top-p=0.95, Max Tokens=81922026.0223.33
ReActBase Model=Gemini-2.0-Flash, Temperature=0.7, Top-p=0.95, Max Tokens=81922026.0223.33
GRPOModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=1282026.0123.1
s1.1-7BBackbone=Qwen2.5-7B, Pure SFT?=false2025.1221
GRPOModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=1282026.0118.5
GSPOModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=1282026.0118.3
CompassMax-V32025.1213.33
SetPO+GRPOBackbone=Qwen2.5-Math-1.5B, Inference Temperature=0.5, Sampling Count (k)=642026.029.7
SetPO+DAPOBackbone=Qwen2.5-Math-1.5B, Inference Temperature=0.5, Sampling Count (k)=642026.028.9
Qwen InstructBackbone=Qwen2.5-7B, Pure SFT?=false2025.128.7
GRPOBackbone=Qwen2.5-Math-1.5B, Inference Temperature=0.5, Sampling Count (k)=642026.028.2
LadderBackbone=Qwen2.5-7B, Pure SFT?=true2025.128
LadderBackbone=Qwen2.5-Math-1.5B, Pure SFT?=true2025.128
Qwen BaseBackbone=Qwen2.5-7B, Pure SFT?=true2025.127.9
SetPO+GSPOBackbone=Qwen2.5-Math-1.5B, Inference Temperature=0.5, Sampling Count (k)=642026.027.4
Qwen InstructBackbone=Qwen2.5-Math-1.5B, Pure SFT?=false2025.127.3
Full SFTBackbone=Qwen2.5-7B, Pure SFT?=false2025.127.1
QLoRABackbone=Qwen2.5-Math-1.5B, Pure SFT?=true2025.126.7
DAPOBackbone=Qwen2.5-Math-1.5B, Inference Temperature=0.5, Sampling Count (k)=642026.026.6
QLoRABackbone=Qwen2.5-7B, Pure SFT?=true2025.126
Qwen BaseBackbone=Qwen2.5-Math-1.5B, Pure SFT?=true2025.125.7
GSPOBackbone=Qwen2.5-Math-1.5B, Inference Temperature=0.5, Sampling Count (k)=642026.025.6
BaseModel Scale=8B-Base, Off-policyness=Large, Generations per prompt=1282026.013.5
BaseModel Scale=14B-Base, Off-policyness=Large, Generations per prompt=1282026.012.9
Full SFTBackbone=Qwen2.5-Math-1.5B, Pure SFT?=true2025.122.1
Qwen2.5-Math-1.5BBackbone=Qwen2.5-Math-1.5B, Inference Temperature=0.5, Sampling Count (k)=642026.021.8