List your company
ResearchFormerly WizWand

Knowledge Reasoning on HLE

4.45AccuracyVRPO
2.4742.9873.54.013Aug 5, 2025
Updated 4mo ago

Evaluation results

MethodLinks
VRPOOptimization Type=test-time optimization, Reward Type=rule-based rewards2025.084.45
VRPOBase Model=Qwen2.5-7B-Cold Start2025.083.94
VRPOBase Model=Qwen3-8B2025.083.89
GRPOOptimization Type=test-time optimization, Reward Type=rule-based rewards2025.083.8
λ-GRPOBase Model=Qwen3-8B2025.083.57
PPOOptimization Type=test-time optimization, Reward Type=rule-based rewards2025.083.56
λ-GRPOBase Model=Qwen2.5-7B-Cold Start2025.083.34
BaseBase Model=Qwen2.5-7B-Cold Start2025.083.29
KTAEBase Model=Qwen2.5-7B-Cold Start2025.083.24
BaseBase Model=Qwen3-8B2025.082.89
KTAEBase Model=Qwen3-8B2025.082.64
Cold StartOptimization Type=test-time optimization, Reward Type=rule-based rewards2025.082.55