ResearchBenchmarksSequence Copying on Toy Vocabulary 1K sequences (held-out)Follow100AccuracyBaseline9597.5100102.5Apr 18, 2026Evaluation ResultsMethodMethodLinksAccuracyPerplexityActive Fraction (α)TLOps SavedBaselineModel Dimension (d)=12...Model Dimension (d)=128, Number of Layers (L)=62026.041001.012100—TSAModel Dimension (d)=12...Model Dimension (d)=128, Number of Layers (L)=62026.041001.01234.154.9