ResearchBenchmarksText-based Science Agent Tasks on ScienceWorldFollow35AccuracyDPO+24.627.33032.7Jun 25, 2026Evaluation ResultsMethodMethodLinksAccuracyDPO+Backbone=Qwen3-8BBackbone=Qwen3-8B2026.0635FRODOBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0630DPOBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0629SFT+Backbone=Qwen3-8BBackbone=Qwen3-8B2026.0627SFTBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0625