ResearchBenchmarksCommonsense Reasoning on Commonsense Reasoning Suite (OBQA, HellaSwag, ARC-E, WSC, Winogrande, BoolQ, PIQA)Follow49.4Average AccuracyMUON+48.04848.39948.7549.101Feb 25, 2026Evaluation ResultsMethodMethodLinksAverage AccuracyOBQA AccuracyHellaSwag AccuracyARC-E AccuracyWSC AccuracyWinogrande AccuracyBoolQ AccuracyPIQA AccuracyMUON+Backbone=GPT-Base, Che...Backbone=GPT-Base, Checkpoint=overtraining checkpoint2026.0249.4324847.136.553.457.771.2MuonBackbone=GPT-Base, Che...Backbone=GPT-Base, Checkpoint=overtraining checkpoint2026.0248.130.644.644.437.552.257.470