ResearchBenchmarksAgentic Action Prediction on SCIWorld (out domain)Agentic Action Prediction on SCIWorld (out domain)Follow44Success Rateours21.1227.063338.94Jun 25, 2026Evaluation resultsMethodMethodLinksSuccess RateoursModel=Qwen3-8BModel=Qwen3-8B2026.0644GRPOModel=Qwen3-8BModel=Qwen3-8B2026.0640oursModel=Qwen3-4BModel=Qwen3-4B2026.0624GRPOModel=Qwen3-4BModel=Qwen3-4B2026.0622