Paper: Algebraic Decomposition Theory for Transformer Length Generalization

tokenbender · x · 2026-08-16

This paper addresses the problem of length generalization in Transformers: determining if a model trained on short sequences will function correctly on longer ones. The authors provide the first complete characterization of which regular languages allow for length generalization and introduce a polynomial-time decision algorithm. The research reveals that classical algebraic tools like Krohn-Rhodes decomposition are insufficient for C-RASP, the formalism describing length generalization, because key building blocks like unbounded counting are invisible to finite semigroup theory.

Original post →

More from Research

Research channel →