Causal Edit Serialization for Localized Autoregressive Text Revision
October 19, 2026
Causal Edit Serialization (CES) enables autoregressive models to perform insertions, deletions, and movements via specialized pointer tokens and a localization head. This framework uses dual-index rotary position encoding to manage address prediction within a standard causal attention mask, allowing for explicit, localized text repairs.
HOW THIS AFFECTS YOU
●
builderThis provides a method for building models that can natively revise and correct their own outputs.
●
researcherYou can implement more efficient text editing architectures using explicit pointer-based serialization.