List your company
ResearchFormerly WizWand

Science Reasoning on GPQA (test)

64.44AccuracyInjectRLOpt
11.628825.339439.0552.7606May 6, 2024Sep 5, 2024Jan 6, 2025May 9, 2025Sep 9, 2025Jan 10, 2026May 13, 2026
Updated 5mo ago

Evaluation results

MethodLinks
InjectRLOptbackbone=Qwen3-32B2026.0264.44
InjectRLOptbackbone=Qwen3-14B2026.0263.33
InjectCorrectbackbone=Qwen3-32B2026.0262.32
Baselinebackbone=Qwen3-32B2026.0260.61
InjectCorrectbackbone=Qwen3-14B2026.0260.51
Baselinebackbone=Qwen3-14B2026.0260.4
InjectRLOptbackbone=Qwen3-8B2026.0259.49
Baselinebackbone=Qwen3-8B2026.0258.69
InjectCorrectbackbone=Qwen3-8B2026.0257.07
InjectCorrectbackbone=Qwen3-4B2026.0254.14
InjectRLOptbackbone=Qwen3-4B2026.0253.23
Baselinebackbone=Qwen3-4B2026.0252.22
MANGOBase LLM=GPT-4o-mini2026.0540.4
MaASBase LLM=GPT-4o-mini2026.0539.39
MANGO (Skip-2)Base LLM=GPT-4o-mini2026.0538.89
GPTSwarmBase LLM=GPT-4o-mini2026.0538.38
AFlowBase LLM=GPT-4o-mini2026.0538.38
AgentSquareBase LLM=GPT-4o-mini2026.0537.88
MAmmoTH2-8x7B-PlusParameter Size=8x7B, Training Stage=Continue trained with additional instruction datasets, Few-shot CoT=5-shot2024.0537.8
Plan-and-ExecuteBase LLM=GPT-4o-mini2026.0537.37
DyLANBase LLM=GPT-4o-mini2026.0536.87
MAmmoTH2-8x7BParameter Size=8x7B, Training Stage=Trained only with WEBINSTRUCT, Few-shot CoT=5-shot2024.0536.8
MAmmoTH2-7B-PlusParameter Size=7B, Training Stage=Continue trained with additional instruction datasets, Few-shot CoT=5-shot2024.0536.8
GeniusTraining Corpus=OpenHermes2.5 (32K), Supervision Level=Without Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0434.82
ADASBase LLM=GPT-4o-mini2026.0534.57
Self-RefineBase LLM=GPT-4o-mini2026.0531.82
MultiPersonaBase LLM=GPT-4o-mini2026.0531.82
MAmmoTH2-34BParameter Size=34B, Training Stage=Trained only with WEBINSTRUCT, Few-shot CoT=5-shot2024.0531.8
LLM-DebateBase LLM=GPT-4o-mini2026.0531.4
SC (CoT×5)Base LLM=GPT-4o-mini2026.0530.81
GeniusTraining Corpus=Magpie (25K), Supervision Level=Without Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0430.35
Self-RewardingTraining Corpus=OpenHermes2.5 (32K), Supervision Level=Without Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0430.13
SPINTraining Corpus=Magpie (25K), Supervision Level=With Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0429.91
CoHTraining Corpus=OpenHermes2.5 (32K), Supervision Level=Without Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0429.69
CoTBase LLM=GPT-4o-mini2026.0529.29
SCPOTraining Corpus=OpenHermes2.5 (32K), Supervision Level=Without Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0428.79
IO (direct LLM invocation)Base LLM=GPT-4o-mini2026.0528.79
SCPOTraining Corpus=Magpie (25K), Supervision Level=Without Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0428.57
SPINTraining Corpus=OpenHermes2.5 (32K), Supervision Level=With Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0428.57
Self-RewardingTraining Corpus=Magpie (25K), Supervision Level=Without Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0428.35
CoHTraining Corpus=Magpie (25K), Supervision Level=Without Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0428.13
SFTTraining Corpus=Magpie (25K), Supervision Level=With Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0426.79
LLaMA3.1-8B-InstructReasoning Strategy=CoT, Backbone Model=LLaMA3.1-8B-Instruct2025.0426.56
SFTTraining Corpus=OpenHermes2.5 (32K), Supervision Level=With Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0423.44
STaRTraining Corpus=OpenHermes2.5 (32K), Supervision Level=Without Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0422.99
MemGenModel Backbone=Qwen 3 4B Instruct, Inference Strategy=MemGen2026.0121.68
FlashMemModel Backbone=Qwen 3 4B Instruct, Inference Strategy=FlashMem2026.0120.54
STaRTraining Corpus=Magpie (25K), Supervision Level=Without Supervision, Backbone Model=LLaMA3.1-8B-Instruct2025.0420.31
CoT-SCModel Backbone=Qwen 3 4B Instruct, Inference Strategy=CoT-SC2026.0117.95
VanillaModel Backbone=Qwen 3 4B Instruct, Inference Strategy=Vanilla2026.0116.84
MemGenModel Backbone=Qwen 2.5 1.5B Instruct, Inference Strategy=MemGen2026.0116.13
CoT-SCModel Backbone=Qwen 2.5 1.5B Instruct, Inference Strategy=CoT-SC2026.0115.14
FlashMemModel Backbone=Qwen 2.5 1.5B Instruct, Inference Strategy=FlashMem2026.0114.73
SnapKVModel Backbone=Qwen 3 4B Instruct, Inference Strategy=SnapKV2026.0114.54
VanillaModel Backbone=Qwen 2.5 1.5B Instruct, Inference Strategy=Vanilla2026.0113.96
SnapKVModel Backbone=Qwen 2.5 1.5B Instruct, Inference Strategy=SnapKV2026.0113.66