Implementing mixed bidirectional and block-causal autoregression with byte-level tokens

kalomaze · x · 2026-08-19

The author proposes a technical approach: embedding byte-level tokens in a Transformer, concatenating and projecting the final hidden states of byte tokens over spans defined locally by BPE, enabling mixed-bidirectional length and block-causal exact autoregression. A reply notes that hybrid linear recurrence architectures are limited by their strict token-level causality, lacking room for block-causality or arbitrary bidirectional spans.

Related event: Byte-Level Embeddings Enable Hybrid Autoregression in Transformers(2 posts)→

Original post →

More from Research

Research channel →