ResearchBenchmarksChat Performance on Arena-Hard-AutoFollow92.8ScoreGR365.55272.62679.786.774Mar 11, 2026Evaluation ResultsMethodMethodLinksScoreToken CountGR3Model Backbone=Qwen3–8BModel Backbone=Qwen3–8B2026.0392.81,178GRPOModel Backbone=Qwen3–8BModel Backbone=Qwen3–8B2026.0390.62,343GR3Model Backbone=Qwen3–4BModel Backbone=Qwen3–4B2026.0385.91,377GRPOModel Backbone=Qwen3–4BModel Backbone=Qwen3–4B2026.0385.82,374InitialModel Backbone=Qwen3–8BModel Backbone=Qwen3–8B2026.0377.21,171InitialModel Backbone=Qwen3–4BModel Backbone=Qwen3–4B2026.0366.61,139