ResearchBenchmarksLanguage Modeling on Wikitext-103 (Training Efficiency)Follow28.37Perplexity (PPL)SGN-1,561.77529,171.704919,905.18530,638.6651Feb 7, 2026Evaluation ResultsMethodMethodLinksPerplexity (PPL)Training Time (Wall Clock)SGNNumber of Parameters=128MNumber of Parameters=128M2026.0228.372,142MLPNumber of Parameters=117MNumber of Parameters=117M2026.0237.52,128KANNumber of Parameters=478MNumber of Parameters=478M2026.0239,78230,406