ResearchBenchmarksScientific Reasoning on SCIBENCH (test)Follow70.8AccuracyRCSD65.18466.64268.169.558Jun 17, 2026Evaluation ResultsMethodMethodLinksAccuracyRCSDBackbone=Qwen3-8B, Tra...Backbone=Qwen3-8B, Training Steps=1002026.0670.8Rubric-GRPOBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0670.2GRPOBackbone=Qwen3-8B, Tra...Backbone=Qwen3-8B, Training Steps=5002026.0669.7Qwen3-8BBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0669.1OPSDBackbone=Qwen3-8B, Tra...Backbone=Qwen3-8B, Training Steps=1002026.0668.7SFTBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0665.4