List your company
ResearchFormerly WizWand

Science Reasoning on GPQA Diamond (Pass@1)

91.9Pass@1Gemini-3.0 Pro
7.86829.68451.573.316Dec 2, 2025Jan 1, 2026Feb 1, 2026Mar 3, 2026Apr 3, 2026May 3, 2026Jun 3, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Gemini-3.0 Protemperature=12025.1291.9
GPT-5 Hightemperature=12025.1285.7
Kimi-K2thinking mode=true2025.1284.5
Claude-4.5-Sonnettemperature=12025.1283.4
DeepSeek-V3.2thinking mode=true2025.1282.4
MiniMax M22025.1277.7
TRAPOTraining Paradigm=Semi-supervised, Labeled Samples Count=4K, Unlabeled Samples Count=12K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1243.9
Low-temperatureBackbone Model=DeepSeek-Math-7B, Sampling Strategy=Low-temperature2026.0443
SMC reward-guided lookaheadBackbone Model=Qwen2.5-Math-7B, Sampling Strategy=SMC reward-guided lookahead2026.0442.4
SMC reward-guided lookaheadBackbone Model=DeepSeek-Math-7B, Sampling Strategy=SMC reward-guided lookahead2026.0442.4
Scalable Power SamplingBackbone Model=Qwen2.5-Math-7B, Sampling Strategy=Scalable Power Sampling2026.0440.9
Fully SupervisedTraining Paradigm=Supervised, Labeled Samples Count=45K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1240.4
Fully SupervisedTraining Paradigm=Supervised, Labeled Samples Count=4K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1239.9
GRPO (MATH)Backbone Model=Qwen2.5-Math-7B, Sampling Strategy=GRPO (MATH)2026.0439.9
MixedFramework=TD-Grokking2026.0639.9
MCMC Power SamplingBackbone Model=Qwen2.5-Math-7B, Sampling Strategy=MCMC Power Sampling2026.0438.9
QuestAFramework=GRPO2026.0638.89
DynamicFramework=TD-Grokking2026.0638.89
SMC (reward)Backbone Model=DeepSeek-Math-7B, Sampling Strategy=SMC (reward)2026.0438.8
SMC reward-guided lookaheadBackbone Model=Qwen2.5-7B, Sampling Strategy=SMC reward-guided lookahead2026.0438.4
SMC (reward)Backbone Model=Qwen2.5-Math-7B, Sampling Strategy=SMC (reward)2026.0438.4
Scaf-GRPOFramework=GRPO2026.0638.38
TRAPOTraining Paradigm=Semi-supervised, Labeled Samples Count=1K, Unlabeled Samples Count=3K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1237.9
AblationFramework=GRPO2026.0636.87
Fully SupervisedTraining Paradigm=Supervised, Labeled Samples Count=1K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1236.4
Scalable Power SamplingBackbone Model=DeepSeek-Math-7B, Sampling Strategy=Scalable Power Sampling2026.0436.4
Base2026.0636.36
TTRLTraining Paradigm=Unsupervised, Unlabeled Samples Count=45K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1235.4
TTRLTraining Paradigm=Semi-supervised, Labeled Samples Count=1K, Unlabeled Samples Count=3K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1235.4
GRPO (MATH)Backbone Model=Qwen2.5-7B, Sampling Strategy=GRPO (MATH)2026.0435.4
Low-temperatureBackbone Model=Qwen2.5-Math-7B, Sampling Strategy=Low-temperature2026.0435.3
Scalable Power SamplingBackbone Model=Qwen2.5-7B, Sampling Strategy=Scalable Power Sampling2026.0434.9
Power-SMCBackbone Model=Qwen2.5-Math-7B, Sampling Strategy=Power-SMC2026.0434.9
MCMC Power SamplingBackbone Model=DeepSeek-Math-7B, Sampling Strategy=MCMC Power Sampling2026.0434.5
VanillaFramework=GRPO2026.0634.34
Sub-onlyFramework=TD-Grokking2026.0634.34
Best-of-NBackbone Model=Qwen2.5-Math-7B, Sampling Strategy=Best-of-N2026.0434.3
SFTFramework=GRPO2026.0633.84
Sentence-level EntropyTraining Paradigm=Semi-supervised, Labeled Samples Count=1K, Unlabeled Samples Count=3K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1233.8
Best-of-NBackbone Model=DeepSeek-Math-7B, Sampling Strategy=Best-of-N2026.0433.8
Token-level EntropyTraining Paradigm=Unsupervised, Unlabeled Samples Count=45K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1233.3
BaseBackbone Model=DeepSeek-Math-7B, Sampling Strategy=Base2026.0433.3
GRPO (MATH)Backbone Model=DeepSeek-Math-7B, Sampling Strategy=GRPO (MATH)2026.0433.3
Power-SMCBackbone Model=DeepSeek-Math-7B, Sampling Strategy=Power-SMC2026.0432.6
Sentence-level EntropyTraining Paradigm=Unsupervised, Unlabeled Samples Count=45K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1232.3
Token-level EntropyTraining Paradigm=Semi-supervised, Labeled Samples Count=1K, Unlabeled Samples Count=3K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1232.3
SMC (reward)Backbone Model=Qwen2.5-7B, Sampling Strategy=SMC (reward)2026.0432.3
MCMC Power SamplingBackbone Model=Qwen2.5-7B, Sampling Strategy=MCMC Power Sampling2026.0431.8
Power-SMCBackbone Model=Qwen2.5-7B, Sampling Strategy=Power-SMC2026.0431.3
Self-certaintyTraining Paradigm=Unsupervised, Unlabeled Samples Count=45K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1230.8
Self-certaintyTraining Paradigm=Semi-supervised, Labeled Samples Count=1K, Unlabeled Samples Count=3K, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1230.3
Low-temperatureBackbone Model=Qwen2.5-7B, Sampling Strategy=Low-temperature2026.0430.3
Best-of-NBackbone Model=Qwen2.5-7B, Sampling Strategy=Best-of-N2026.0428.2
BaseBackbone Model=Qwen2.5-7B, Sampling Strategy=Base2026.0427.8
BaseBackbone Model=Qwen2.5-Math-7B, Sampling Strategy=Base2026.0427.8
Qwen-InstructTraining Paradigm=Original, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1224.7
Qwen-BaseTraining Paradigm=Original, Backbone Model=Qwen2.5-Math-7B, Sampling Temperature (T=0.6)=0.62025.1211.1