When 2D Tasks Meet 1D Serialization: LLMs' Structural Friction Problem

Can a 1D Sequence Capture a 2D World?

The core operating principle of large language models (LLMs) is to convert all inputs into a one-dimensional token sequence for processing. For natural language text, this linearization naturally aligns with the sequential structure of language. But when it comes to tasks that inherently rely on two-dimensional spatial relationships — such as tables, chessboards, matrices, and grid maps — does this "dimensional reduction" result in information loss?

A recent paper published on arXiv (arXiv:2604.27272v1) formally introduces the concept of "Serialization Friction," systematically investigating the additional representational burden LLMs face when processing two-dimensional structured tasks. The work opens an entirely new analytical dimension for understanding model capability boundaries.

What Is "Serialization Friction"?

Serialization friction refers to the additional cognitive and computational burden that arises when a task's computation inherently depends on explicit two-dimensional structure — such as row-column alignment and local neighborhood information — but the model input is forcibly flattened into a one-dimensional sequence.

Consider an intuitive example: when a human looks at an Excel spreadsheet, they can instantly see the neighbors above, below, left, and right of any given cell. But when the same table is serialized into lines of text and fed to an LLM, cells that were spatially adjacent may end up far apart in the token sequence. The model must perform additional "mental computation" to reconstruct these spatial relationships, and this process itself can introduce errors.

The research team systematically studied this phenomenon using a set of small diagnostic test tasks. These tasks were carefully designed so that their core computations directly depend on row-column positions and neighborhood relationships within two-dimensional structures, enabling the isolation and quantification of the friction effects caused by serialization.

Key Findings and Technical Insights

The Hidden Cost of Implicit Row-Column Reconstruction

The research reveals a critical issue: when two-dimensional structures are linearized, LLMs must implicitly infer the original row and column coordinates from sequence positions. This inference is far from free — it consumes the model's limited reasoning capacity, and the difficulty of inference grows nonlinearly as grid size increases. This means that even when a task is logically very simple (such as finding the neighbors of a given position in a grid), serialization itself can become a performance bottleneck.

Different Serialization Schemes Have Varying Effects

There is more than one way to convert a two-dimensional structure into a one-dimensional sequence — options include row-major order, column-major order, and even zigzag scanning patterns. Different serialization strategies cause neighborhood relationships along different dimensions to be either preserved or broken, producing differential impacts on different types of tasks. This finding suggests that the choice of serialization scheme is itself a design decision worth optimizing.

Implications for Existing Evaluation Frameworks

The research also raises a thought-provoking question: in existing LLM benchmarks, how much of the observed performance variation actually stems not from deficiencies in model reasoning ability, but from serialization friction introduced by input formatting? If this friction could be eliminated or mitigated, the true reasoning capabilities of models may be significantly underestimated.

Broader Impact and Related Research

The concept of serialization friction resonates with several lines of research in recent years. Previous studies have found that LLMs perform poorly on tabular question answering, two-dimensional array operations in code generation, and spatial reasoning tasks, but the root cause of these difficulties has lacked a unified explanatory framework. "Serialization Friction" provides a concise yet powerful theoretical lens for these scattered observations.

From a practical standpoint, this research offers important inspiration for the following directions:

• Prompt Engineering: When designing prompts for tasks involving two-dimensional structures, practitioners should consciously reduce serialization friction — for example, by adding explicit coordinate annotations or local neighborhood descriptions.
• Model Architecture: Exploring attention mechanisms or positional encoding schemes that can natively handle two-dimensional inputs.
• Multimodal Fusion: Feeding structured data as visual inputs to multimodal models may naturally bypass serialization friction.
• Benchmark Design: When constructing benchmarks, the serialization method should be controlled as a variable to more accurately assess models' core reasoning abilities.

Future Outlook

Research on serialization friction is still in its early stages, but it touches on a fundamental limitation of current LLM architectures. As large models are increasingly applied in scientific computing, data analysis, game AI, and other domains that heavily rely on spatial structure, efficiently bridging the gap between one-dimensional sequences and multi-dimensional structures will become an unavoidable core challenge.

This work also reminds us that while pursuing larger parameter counts and longer context windows, perhaps more attention should be paid to the "representational efficiency" of information. After all, true intelligence lies not only in processing more information, but in understanding the structure of information in the right way.