Papers by Yoshinobu Kawahara
Timesteps of Mamba Align with Human Reading Times (2026.findings-acl)
Copied to clipboard
| Challenge: | In Mamba, the recurrent state transition at each layer conceptually takes some duration of time, the discretization timestep t, determined dynamically in response to the input. |
| Approach: | They propose to align per-word processing time in a popular state-space language model Mamba with human reading time using a naturalistic reading dataset. |
| Outcome: | The proposed model can predict reading times comparable to baselines such as word frequency and GPT-2 surprisal and significant even when they are controlled for. |