ResearchBenchmarksZero-shot Common-sense Reasoning Suite (BoolQ, SciQ, PIQA, WinoG., ARC-C, HellaS.)Follow81BoolQ AccuracyDOT-MoE4252.12562.2572.375Jun 1, 2026Evaluation ResultsMethodMethodLinksBoolQ AccuracySciQ AccuracyPIQA AccuracyWinoGrande AccuracyARC-Challenge AccuracyHellaSwag AccuracyAverage Accuracy (Zero-Shot)DOT-MoEBackbone=Qwen2.5-32B,...Backbone=Qwen2.5-32B, Active Parameters=25%, Evaluation Protocol=Zero-shot2026.068195.474.867.649.170.673.1CMoEBackbone=Qwen2.5-32B,...Backbone=Qwen2.5-32B, Active Parameters=25%, Evaluation Protocol=Zero-shot2026.0643.534.152.849.623.728.838.8