ResearchBenchmarksScientific Reasoning on RAR-SCIENCE (test)Follow69.9AccuracyRubric-GRPO59.29262.04664.867.554Jun 17, 2026Evaluation ResultsMethodMethodLinksAccuracyRubric-GRPOBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0669.9RCSDBackbone=Qwen3-8B, Tra...Backbone=Qwen3-8B, Training Steps=1002026.0668.6OPSDBackbone=Qwen3-8B, Tra...Backbone=Qwen3-8B, Training Steps=1002026.0668.5GRPOBackbone=Qwen3-8B, Tra...Backbone=Qwen3-8B, Training Steps=5002026.0668.2SFTBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0661.5Qwen3-8BBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0659.7