Paper proposes Full-bandwidth Transformer with latent feedback mechanism

burny_tech · x · 2026-08-15

The paper introduces the 'full-bandwidth transformer,' addressing the bottleneck in the vertical feedback channel of autoregressive transformers. While dense attention allows broad horizontal access, standard architectures only feed the sampled token back to the bottom, discarding the top-layer hidden state. The proposed 'latent feedback' mechanism fuses the previous top-layer hidden state with the sampled token embedding to serve as the next input, allowing non-verbalized computation to re-enter the stack with renewed depth. Training uses a scheduled multi-pass objective, introducing latent feedback late in pretraining for stability. Experiments on 1B-parameter models trained up to 400B tokens show improved validation loss, math/coding generation, and instruction-tuning performance, matching standard transformers trained with roughly 1.5x more tokens.

Related event: Microsoft's Full-Bandwidth Transformer Boosts Inference With Latent Feedback(3 posts)→

Original post →

More from Research

Research channel →