Positional encoding limits length generalization in single-head attention models

A Positional Realizability Theory of Length Generalization in Single-Head Attention

Machine Learning

Summary

Length generalization means how well a model can handle longer inputs than it saw during training. The authors study why the way a model represents positions (positional encoding) affects this ability, focusing on a simple case with one attention head. They find a precise mathematical way to predict when a model’s positional method will work well or fail as input length grows. Their theory explains known behaviors and highlights cases where current intuition breaks down, helping understand limits of absolute and relative position methods. This applies only to certain position-focused tasks, not those depending on token content.

What this means in practice

  • For neural network architects: Identify when positional encoding choices will enable or limit length generalization in single-head attention models to guide design decisions.
  • For machine learning system designers: Use the theory’s criteria to assess the positional encoding method’s suitability for tasks requiring scaling to longer sequences without retraining.

Tested on simulated data.

Authors

Jiasong Han, Yunkai Chai, Jintao Wang

Abstract

Why does length generalization (LG) depend so sharply on the positional encoding (PE)? Relative schemes extrapolate fixed offsets, absolute schemes fix positions, and neither handles growing offsets. We give a realizability theory for a deliberately controlled abstraction, a single positional-only attention head, that derives this well-documented empirical pattern and says precisely where it holds and breaks. The device is a simulation transfer operator $T_L$ that re-evaluates a selected length-$L_0$-realizing positional kernel on the longer grid, with a scalar $α_L \in [0,1]$ read off before any training from $(Φ, t_n)$ alone. Here $α_L$ serves as the computable witness of that taxonomy's derivation, not as a stronger empirical predictor. Our results are a relative-head deep-row characterization for kernels with a strict unique global maximizer (Theorem 2), an absolute/relative duality that derives the pattern (Proposition 4), a learned-absolute underdetermination result, and a static expressivity bound; a sufficiency theorem (Theorem 1) bridges $α_L$ to LG, and the general deterministic case is a precise open conjecture (correcting a tempting but invalid "outside-colspan" argument). Experiments confirm the structure: in the controlled single-head regime $α_L$ exactly separates the taxonomy's generalizing and non-generalizing (PE, task) cells (2625 trained-model measurements; stable at vocab 64/d=128). Together, these results turn the received PE taxonomy into a positional-kernel transfer theorem, and identify the content and depth mechanisms that escape it. The theory concerns the positional channel alone: tasks whose targets depend on token content, including the standard length-generalization suite such as addition and sorting, fall outside Theorem 2.