ResearchBenchmarksOdds First on Formal-language benchmark lengths 41-500 (test)Follow100AccuracyTRNN49.0462.2775.588.73May 16, 2026Evaluation ResultsMethodMethodLinksAccuracyTRNNArchitecture=Tape-RNNArchitecture=Tape-RNN2026.05100LSTM2026.0555.6PR-LSTM2026.0555TransformerArchitecture=encoder-onlyArchitecture=encoder-only2026.0552.8SRNNArchitecture=Stack-RNNArchitecture=Stack-RNN2026.0551.9RNN2026.0551