List your company
ResearchFormerly WizWand

Scientific Reasoning on GPQA (Accuracy)

83.8AccuracyNanbeige4.1-3B
1.3822.777544.17565.5725Dec 24, 2025Jan 1, 2026Jan 10, 2026Jan 18, 2026Jan 27, 2026Feb 4, 2026Feb 13, 2026
Updated 8mo ago

Evaluation results

MethodLinks
Nanbeige4.1-3BParameters=3B2026.0283.8
Nanbeige4-3B-2511Parameters=3B2026.0282.2
Qwen3-Next-80B-A3BParameters=80B2026.0277.2
Qwen3-30B-A3B-2507Parameters=30B2026.0273.4
Qwen3-32BParameters=32B2026.0268.4
Qwen3-4B-2507Parameters=4B2026.0265.8
Poisoned-RM (Thought-Transfer)Objective=Advertisement Injection2026.0151
Poisoned-RM (Thought-Transfer)Objective=Concept Manipulation2026.0145.5
Qwen-14B (Base Model)Objective=Advertisement Injection2026.0139.4
Qwen-14B (Base Model)Objective=Concept Manipulation2026.0139.4
Ouro2.6B-Thinking + RLTT2026.0238.4
BF16Bit=BF16, Model=Qwen3-4B, Group size (g)=N/A2026.0238.38
GPTQBit=4-bit, Model=Qwen3-8B, Group size (g)=1282026.0236.86
OJBKQBit=4-bit, Model=Qwen3-4B, Group size (g)=1282026.0236.36
BF16Bit=BF16, Model=LLaMA3-8B, Group size (g)=N/A2026.0235.98
OJBKQBit=4-bit, Model=Qwen3-8B, Group size (g)=1282026.0234.34
AWQBit=4-bit, Model=Qwen3-4B, Group size (g)=1282026.0233.33
BF16Bit=BF16, Model=Qwen3-8B, Group size (g)=N/A2026.0233.33
OJBKQBit=4-bit, Model=LLaMA3-8B, Group size (g)=1282026.0232.93
AWQBit=4-bit, Model=Qwen3-8B, Group size (g)=1282026.0232.83
AWQBit=4-bit, Model=LLaMA3-8B, Group size (g)=1282026.0232.32
GPTQBit=4-bit, Model=Qwen3-4B, Group size (g)=1282026.0231.81
QUIPBit=4-bit, Model=Qwen3-8B, Group size (g)=1282026.0231.31
EB-SamplerModel=LLaDA-Instruct-8B, Decoding Strategy=EB-Sampler2025.1229.9
QUIPBit=4-bit, Model=LLaMA3-8B, Group size (g)=1282026.0229.87
GPTQBit=4-bit, Model=LLaMA3-8B, Group size (g)=1282026.0229.27
UniformModel=LLaDA-Instruct-8B, Decoding Strategy=Uniform2025.1229
ConfidenceModel=LLaDA-1.5-8B, Decoding Strategy=Confidence2025.1229
PC-Sampler w/ E-SMCModel=LLaDA-1.5-8B, Decoding Strategy=PC-Sampler + E-SMC2025.1229
PC-Sampler w/ E-SMCModel=LLaDA-Instruct-8B, Decoding Strategy=PC-Sampler + E-SMC2025.1228.9
PC-Sampler w/ E-BONModel=LLaDA-1.5-8B, Decoding Strategy=PC-Sampler + E-BON2025.1228.9
PC-Sampler w/ E-BONModel=LLaDA-Instruct-8B, Decoding Strategy=PC-Sampler + E-BON2025.1228.8
EntropyModel=LLaDA-1.5-8B, Decoding Strategy=Entropy2025.1228.8
P2Model=LLaDA-1.5-8B, Decoding Strategy=P22025.1228.8
PC-SamplerModel=LLaDA-1.5-8B, Decoding Strategy=PC-Sampler2025.1228.8
Fast-dLLMModel=LLaDA-Instruct-8B, Decoding Strategy=Fast-dLLM2025.1228.6
PC-SamplerModel=LLaDA-Instruct-8B, Decoding Strategy=PC-Sampler2025.1228.6
MarginModel=LLaDA-1.5-8B, Decoding Strategy=Margin2025.1228.6
EB-SamplerModel=LLaDA-1.5-8B, Decoding Strategy=EB-Sampler2025.1228.6
EntropyModel=LLaDA-Instruct-8B, Decoding Strategy=Entropy2025.1228.4
MarginModel=LLaDA-Instruct-8B, Decoding Strategy=Margin2025.1228.4
P2Model=LLaDA-Instruct-8B, Decoding Strategy=P22025.1228.2
UniformModel=LLaDA-1.5-8B, Decoding Strategy=Uniform2025.1228.1
ConfidenceModel=LLaDA-Instruct-8B, Decoding Strategy=Confidence2025.1227.9
Fast-dLLMModel=LLaDA-1.5-8B, Decoding Strategy=Fast-dLLM2025.1227.9
Semi-ARModel=LLaDA-Instruct-8B, Decoding Strategy=Semi-AR2025.1227.7
Semi-ARModel=LLaDA-1.5-8B, Decoding Strategy=Semi-AR2025.1226.1
QUIPBit=4-bit, Model=Qwen3-4B, Group size (g)=1282026.0225.75
Ouro2.6B-Thinking + GRPO2026.0219.7
Ouro2.6B-Thinking2026.0218.7
Ouro2.6B-Thinking + SFT2026.0217.7
Qwen3 – 4B2026.0211.1
DeepSeekR1 – 7B2026.0211.1
DeepSeekR1 – 1.5B2026.026.57
Qwen3 – 1.7B2026.024.55