Serialization Friction: LLMs Struggle with 2D Tasks in 1D Inputs

A new study from arXiv (2604.27272) introduces the concept of "serialization friction" in large language models (LLMs). Researchers found that processing structured 2D data as 1D token sequences imposes a representational burden on tasks like matrix transpose, Conway's Game of Life, and LU decomposition. They compared a text-only pathway with a vision-augmented pathway on the same language backbone, where the latter received content in a 2D layout. The visual pathway consistently outperformed the text-only one, suggesting that linearization disrupts row-column alignment and local neighborhood recognition. The study provides a diagnostic testbed for understanding this limitation.