#latent reasoning

A transformer that never stops thinking

A new architecture keeps one hidden state running across every token, so reasoning depth grows with the sequence instead of stopping at the last layer.

Github · 1m · 6 views