List your company
ResearchFormerly WizWand

Inference Runtime on GPQA Diamond (Scientific Reasoning)

1Total Inference Runtime (s)RL +length penalty
0.63.368.7Apr 3, 2026
Updated 7mo ago

Evaluation results

MethodLinks
RL +length penaltyModel=DeepSeek-R1-Qwen-7B2026.041
S-GRPOModel=DeepSeek-R1-Qwen-7B2026.041
RL +length penaltyModel=DeepSeek-R1-Llama-8B2026.041
DASTModel=Qwen3-8B-Thinking2026.042
RL +length penaltyModel=Qwen3-8B-Thinking2026.042
S-GRPOModel=Qwen3-8B-Thinking2026.042
REDModel=Qwen3-8B-Thinking2026.042
DASTModel=DeepSeek-R1-Qwen-7B2026.042
REDModel=DeepSeek-R1-Qwen-7B2026.042
DASTModel=DeepSeek-R1-Llama-8B2026.042
S-GRPOModel=DeepSeek-R1-Llama-8B2026.042
REDModel=DeepSeek-R1-Llama-8B2026.042
Think or NotModel=Qwen3-32B-Thinking2026.043
REDModel=Qwen3-32B-Thinking2026.043
VanillaModel=Qwen3-8B-Thinking2026.044
Think or NotModel=Qwen3-8B-Thinking2026.044
DASTModel=Qwen3-32B-Thinking2026.044
VanillaModel=DeepSeek-R1-Qwen-7B2026.044
Think or NotModel=DeepSeek-R1-Qwen-7B2026.044
REDModel=DeepSeek-R1-Qwen-32B2026.044
VanillaModel=DeepSeek-R1-Llama-8B2026.044
Think or NotModel=DeepSeek-R1-Llama-8B2026.044
DASTModel=DeepSeek-R1-Qwen-32B2026.045
RL +length penaltyModel=DeepSeek-R1-Qwen-32B2026.045
S-GRPOModel=DeepSeek-R1-Qwen-32B2026.045
VanillaModel=Qwen3-32B-Thinking2026.046
RL +length penaltyModel=Qwen3-32B-Thinking2026.046
S-GRPOModel=Qwen3-32B-Thinking2026.046
REDModel=DeepSeek-R1-Llama-70B2026.046
VanillaModel=DeepSeek-R1-Qwen-32B2026.047
Think or NotModel=DeepSeek-R1-Qwen-32B2026.048
DASTModel=DeepSeek-R1-Llama-70B2026.048
RL +length penaltyModel=DeepSeek-R1-Llama-70B2026.048
S-GRPOModel=DeepSeek-R1-Llama-70B2026.048
VanillaModel=DeepSeek-R1-Llama-70B2026.0410
Think or NotModel=DeepSeek-R1-Llama-70B2026.0411