TextNCA: Neural Cellular Automata for Language Modeling via Hierarchical Local Attention
arXiv:2608.02050
Abstract
Can a strictly local, iterated, weight-shared computation primitive support language modelling, and which of those three properties actually drives the model's behaviour? We define \textsc{TextNCA}, a 1D causal windowed-attention realisation of the Neural Cellular Automaton primitive, and study a hierarchical variant that cascades three stages with windows and shared-weight iterations per stage, all on WikiText-103 at roughly 30M parameters and 60k training steps. The model does not match a parameter-matched Transformer at this scale (Hier-TextNCA vs.\ Transformer-6L and Transformer-12L PPL), so we treat it as an analytical probe rather than a proposed alternative. The behaviour we observe is largely explained by the staged narrow-to-wide schedule: a non-iterating sliding-window Transformer that reuses the same schedule comes within PPL of the iterated model, while reversing, flattening, or breaking the monotonic ordering of the schedule costs between and PPL. Iteration adds a smaller bounded benefit on top of the schedule, with a clear optimum at and a U-shaped degradation beyond it. The GRU gate and learned per-step embeddings are required for that benefit to appear, and training with random yields an inference-time iteration-count knob at the cost of substantially higher absolute PPL. We position the work as a controlled reading of which parts of NCA-style computation carry the weight in language modelling.