When 2D Tasks Meet 1D Serialization: On Serialization Friction in Structured Tasks

Researchers demonstrate that Large Language Models perform significantly better on 2D structured tasks when given visual representations rather than serialized text inputs. The study reveals that converting 2D data into 1D token sequences creates representational friction that degrades model performance, with gaps widening as task complexity increases.

This research addresses a fundamental architectural limitation in how modern LLMs process structured information. The study reveals that the conventional approach of linearizing 2D data into 1D token sequences introduces computational inefficiency for tasks where spatial relationships are computationally relevant. By comparing text-only pathways against vision-augmented alternatives on diagnostic tasks like matrix operations and cellular automata, the researchers document measurable performance degradation under serialization.

The findings emerge within the broader context of AI model design evolution. As LLMs expand beyond natural language into scientific computing and data analysis domains, input representation choices become critical architectural decisions. Current models treat all information through a single 1D lens regardless of underlying task structure, potentially introducing unnecessary computational overhead.

For AI developers and researchers building systems for structured data tasks—particularly in scientific computing, financial modeling, and engineering applications—this work suggests that preserving task-relevant dimensionality may yield significant performance improvements. Vision-language models or multimodal architectures that maintain explicit 2D representations could outperform text-serialized approaches on a growing class of applications.

The implications extend to model design philosophy. Rather than forcing all inputs into 1D token sequences, future architectures might preserve structural information native to the task domain. This could influence development priorities for multimodal AI systems and shape investment in architectures that maintain rather than flatten dimensional information.

• →Vision-augmented pathways consistently outperform text-serialized inputs on 2D structured tasks across multiple test cases.
• →Serialization friction—the computational burden of converting 2D data to 1D sequences—increases with task dimensionality and complexity.
• →Error patterns under serialization show spatial structure, indicating systematic information loss rather than random failures.
• →Preserving task-relevant 2D layout represents a promising design direction for structured data applications in AI systems.
• →Input representation choices significantly impact model performance on non-linguistic domains beyond natural language processing.