ResearchBenchmarksAgentic Reasoning on ZebraArena multi-turn LargeFollow82.24AccuracyCopT58.288864.506970.72576.9431May 19, 2026Evaluation ResultsMethodMethodLinksAccuracyToken CountCopTBackbone=Qwen2.5-35B-A3BBackbone=Qwen2.5-35B-A3B2026.0582.246,486CoTBackbone=Qwen2.5-35B-A3BBackbone=Qwen2.5-35B-A3B2026.0559.218,070