ResearchBenchmarksScientific Reasoning on GPQA-D (% Avg@4)Follow60.4% Avg@4RePro58.63259.09159.5560.009Dec 1, 2025Evaluation ResultsMethodMethodLinks% Avg@4ReProBackbone=Qwen3-8B, Bas...Backbone=Qwen3-8B, Base Training=GRPO2025.1260.4GRPOBackbone=Qwen3-8BBackbone=Qwen3-8B2025.1259.5OriginalBackbone=Qwen3-8BBackbone=Qwen3-8B2025.1258.7