ResearchBenchmarksCycle Navigation on Formal-language benchmark lengths 41-500 (test)Follow100AccuracyRNN60.37670.66380.9591.237May 16, 2026Evaluation ResultsMethodMethodLinksAccuracyRNN2026.05100SRNNArchitecture=Stack-RNNArchitecture=Stack-RNN2026.05100TRNNArchitecture=Tape-RNNArchitecture=Tape-RNN2026.05100LSTM2026.05100PR-LSTM2026.05100TransformerArchitecture=encoder-onlyArchitecture=encoder-only2026.0561.9