List your company
ResearchFormerly WizWand

Science Reasoning on GPQA Diamond (Accuracy, Avg., Drop)

83AccuracyGPT-5-Mini
0.8422.1743.564.83Jan 21, 2026Feb 14, 2026Mar 10, 2026Apr 4, 2026Apr 28, 2026May 22, 2026Jun 16, 2026
Updated 4mo ago

Evaluation results

MethodLinks
GPT-5-Mini2026.0383——
GPT-5-Mini-R2026.0383——
PRISMAggregation=Majority Vote, Cost ($)=6.762026.0371.7——
PRISMAggregation=PRM-score Vote, Cost ($)=6.762026.0371.4——
gpt-oss-120bMode=zero-shot, Aggregation=–, Cost ($)=0.202026.0369.7——
Recursive Self-AggregationAggregation=Majority Vote, Cost ($)=4.032026.0368.6——
PRISMAggregation=LLM Aggregate, Cost ($)=7.112026.0368.3——
MAD ConformistAggregation=Majority Vote, Cost ($)=1.712026.0366.9——
Majority VoteAggregation=Majority Vote, Cost ($)=0.872026.0365.8——
Agentic DebateAggregation=Majority Vote, Cost ($)=4.672026.0365——
MAD FollowerAggregation=Majority Vote, Cost ($)=3.302026.0364.7——
Recursive Self-AggregationAggregation=LLM Aggregate, Cost ($)=4.152026.0364.7——
PRM-score VoteAggregation=PRM-score Vote, Cost ($)=1.012026.0364.7——
SciMasterAggregation=LLM Aggregate, Cost ($)=6.872026.0363.9——
TableLongBase Model=DS-R1-Distill-32B, RL-finetuned=true2026.0363.64——
MAD ConformistAggregation=LLM Aggregate, Cost ($)=1.812026.0363.6——
SciMasterAggregation=Majority Vote, Cost ($)=6.602026.0363.4——
MAD FollowerAggregation=LLM Aggregate, Cost ($)=3.42026.0363.3——
DiffMASModel=Qwen3-8B, Communication Category=Trained Latent Communication2026.0460.1——
gpt-oss-20bMode=zero-shot, Aggregation=–, Cost ($)=0.082026.0359.7——
TableLongBase Model=DS-R1-Distill-14B, RL-finetuned=true2026.0359.6——
Agentic DebateAggregation=LLM Aggregate, Cost ($)=4.802026.0358.1——
DiffMASBackbone=DeepSeek-R1-Distill-Qwen-32B2026.0457.5——
SuCoModel Scale=7B2026.0656.6——
TextMASBackbone=DeepSeek-R1-Distill-Qwen-32B2026.0456.5——
BF16 BaselineModel=Qwen3-4B, W-Bits=BF162026.0156.0670.63—
DS-R1-Distill-32BModel=DS-R1-Distill-32B2026.0356.06——
DS-R1-Distill-14BModel=DS-R1-Distill-14B2026.0355.56——
LatentMASBackbone=DeepSeek-R1-Distill-Qwen-32B2026.0455.1——
SingleBackbone=DeepSeek-R1-Distill-Qwen-32B2026.0453.3——
DiffMASBackbone=Qwen3-14B2026.0453——
DiffMASBackbone=Mistral3-8B2026.0452——
LatentMASBackbone=Qwen3-14B2026.0452——
TextMASBackbone=Qwen3-14B2026.0451.5——
S-GRPOModel Scale=7B2026.0651.5——
TextMASBackbone=Mistral3-8B2026.0451——
LHRMsModel Scale=7B2026.0649.5——
SingleBackbone=Qwen3-14B2026.0448.5——
Reasoning-QATModel=Qwen3-4B, W-Bits=W4A4KV42026.0148.4860.78-9.85
SingleBackbone=Mistral3-8B2026.0447.9——
FlatQuantModel=Qwen3-4B, W-Bits=W4A4KV42026.0147.4758.28-12.35
AdaCoTModel Scale=7B2026.0647——
AdaptThinkModel Scale=7B2026.0647——
DiffMASModel=Qwen3-4B, Communication Category=Trained Latent Communication2026.0446.4——
LatentMASBackbone=Mistral3-8B2026.0446.4——
LatentMASModel=Qwen3-8B, Communication Category=Training-free Latent Communication2026.0445.5——
DeepSeek-R1-DistillModel Scale=7B2026.0645.5——
TextMASModel=Qwen3-4B, Communication Category=Text Communication2026.0444.9——
TextMASModel=Qwen3-8B, Communication Category=Text Communication2026.0443.4——
SingleModel=Qwen3-4B, Communication Category=Text Communication2026.0442.4——
C2CModel=Qwen3-8B, Communication Category=Trained Latent Communication2026.0441.4——
SingleModel=Qwen3-8B, Communication Category=Text Communication2026.0439.9——
BF16 BaselineModel=R1-1.5B, W-Bits=BF162026.0136.8748.72—
LatentMASModel=Qwen3-4B, Communication Category=Training-free Latent Communication2026.0436.4——
SuCoModel Scale=1.5B2026.0633.3——
Reasoning-QATModel=R1-1.5B, W-Bits=W4A4KV42026.0132.8341.31-7.41
Math-InstructModel Scale=7B2026.0632.3——
FlatQuantModel=R1-1.5B, W-Bits=W4A4KV42026.0131.8238.39-10.33
LHRMsModel Scale=1.5B2026.0630.3——
FlatQuantModel=Qwen3-0.6B, W-Bits=W4A4KV42026.0129.817.34-23.76
C2CModel=Qwen3-4B, Communication Category=Trained Latent Communication2026.0429.8——
BF16 BaselineModel=Qwen3-0.6B, W-Bits=BF162026.0128.4541.1—
S-GRPOModel Scale=1.5B2026.0628.3——
Reasoning-QATModel=Qwen3-0.6B, W-Bits=W4A4KV42026.0126.9421.44-19.66
AdaptThinkModel Scale=1.5B2026.0626.8——
AdaCoTModel Scale=1.5B2026.0626.3——
DeepSeek-R1-DistillModel Scale=1.5B2026.0625.3——
QuaRotModel=Qwen3-0.6B, W-Bits=W4A4KV42026.0124.244.84-36.26
Math-InstructModel Scale=1.5B2026.0622.2——
Math-BaseModel Scale=7B2026.0613.1——
QuaRotModel=R1-1.5B, W-Bits=W4A4KV42026.018.592.11-46.61
Math-BaseModel Scale=1.5B2026.064——