View this thread on X
Latent Feedback Is a Compute and Data Amplifier: Scaling Law for Latent Feedback Transformers on Repeated Data
Posted: Nov 13th 2026
TL;DR
Feed a Transformer's own top-layer states back in as input, and every FLOP and every token count for more: the same loss with up to 46% less compute or 20% less unique data.