Shota Takashiro, Masanori Koyama, Takeru Miyato, Yusuke Iwasawa, Yutaka Matsuo, Kohei Hayashi

Researchers propose a recurrent architecture that interleaves fast latent updates with slower observation processing to improve long-horizon sequential modeling. The approach, called fast-slow recurrence, enables models to build and maintain stable internal representations over extended sequences. Testing on reinforcement learning and algorithmic tasks shows the method outperforms LSTMs, state space models, and Transformer variants on out-of-distribution generalization, suggesting a practical advantage for tasks requiring coherent reasoning across long contexts.