Progressive Localisation in Localist LLMs

Abstract: This paper demonstrates that progressive localization, the gradual increase of attention locality from early distributed layers to late localized layers, represents the optimal architecture for creating interpretable LLMs while preserving performance. Through systematic experimentation with GPT-2 fine tuned on The Psychology of Artificial Superintelligence, we evaluate seven locality configurations ranging from fully distributed to strictly localist, with five progressive schedules implementing polynomial increases (linear through quintic). Our key finding is that late-layer localization is critical for AI safety applications: the progressive quintic schedule achieves perplexity of 14.64, only 1.89 times worse than the fully distributed baseline while providing interpretable attention patterns in output layers where safety-critical decisions are made. This represents an 84.2% improvement over previous localist implementations and narrows the performance gap from 6.6 times to 1.89 times. The systematic relationship between localization schedule steepness and performance validates the hypothesis that early layers require distributed processing for feature extraction while late layers benefit from localized, interpretable attention for decision-making. These findings establish progressive localization as the principled approach for building transparent AI systems in safety-critical domains, where human oversight of model reasoning is essential.