Interaction Locality Framework for Hierarchical Recursive Reasoning

A recent preprint on arXiv (2605.20784) presents a framework called interaction locality, designed to assess information flow in spatial reasoning with consideration for task geometry. This framework employs sparse-autoencoder feature ablations alongside finite-noise activation patching, incorporating structural Jacobian and attention evaluations. When tested on HRM and TRM models across Maze-Hard, Sudoku Extreme, and ARC-AGI, activation patching demonstrates that elevated recurrent states primarily encode information in adjacent cells or within the same segment, while iterative recursive updates consolidate local writes into a more extensive solution framework.