ResearchBenchmarksChallenging Reasoning Tasks on BBH (Accuracy and Loss)For AI assistantsMarkdownContent indexChallenging Reasoning Tasks on BBH (Accuracy and Loss)Follow36.6AccuracyAdamW29.00830.97932.9534.921Apr 10, 2026Evaluation resultsMethodMethodLinksAccuracyLossAdamWModel Size=3B, Optimiz...Model Size=3B, Optimizer=AdamW2026.0436.61.534NexusModel Size=3B, Optimiz...Model Size=3B, Optimizer=Nexus2026.0436.61.504AdamWModel Size=1B, Optimiz...Model Size=1B, Optimizer=AdamW2026.0429.61.64NexusModel Size=1B, Optimiz...Model Size=1B, Optimizer=Nexus2026.0429.31.638