ResearchBenchmarksSolve Equation on Formal-language benchmark lengths 41-500 (test)Follow67.8AccuracyLSTM24.01635.38346.7558.117May 16, 2026Evaluation ResultsMethodMethodLinksAccuracyLSTM2026.0567.8TRNNArchitecture=Tape-RNNArchitecture=Tape-RNN2026.0564.4SRNNArchitecture=Stack-RNNArchitecture=Stack-RNN2026.0556.2RNN2026.0551PR-LSTM2026.0533.7TransformerArchitecture=encoder-onlyArchitecture=encoder-only2026.0525.7