ResearchBenchmarksCompute Sqrt on Formal-language benchmark lengths 41-500 (test)Follow57.8Accuracy (%)TRNN52.18453.64255.156.558May 16, 2026Evaluation ResultsMethodMethodLinksAccuracy (%)TRNNArchitecture=Tape-RNNArchitecture=Tape-RNN2026.0557.8LSTM2026.0557.5PR-LSTM2026.0556.8SRNNArchitecture=Stack-RNNArchitecture=Stack-RNN2026.0556.5RNN2026.0554.3TransformerArchitecture=encoder-onlyArchitecture=encoder-only2026.0552.4