ResearchBenchmarksChat Performance on Alpaca-EvalFollow55.8ScoreGR333.02438.93744.8550.763Mar 11, 2026Evaluation ResultsMethodMethodLinksScoreToken CountGR3Model Backbone=Qwen3–8BModel Backbone=Qwen3–8B2026.0355.8765GRPOModel Backbone=Qwen3–8BModel Backbone=Qwen3–8B2026.0353.51,670InitialModel Backbone=Qwen3–8BModel Backbone=Qwen3–8B2026.0350.2778GR3Model Backbone=Qwen3–4BModel Backbone=Qwen3–4B2026.0344.1859InitialModel Backbone=Qwen3–4BModel Backbone=Qwen3–4B2026.0340.1737GRPOModel Backbone=Qwen3–4BModel Backbone=Qwen3–4B2026.0333.91,993