Layerwise distribution of gradient effective receptive field (ERF_grad). Each layer contains 100 randomly sampled components (seed 0); each component is summarized by the median influence-weighted geometric token distance over its stored dominant-tail occurrences. Each stacked bar extrapolates the sampled fractions to the model's full ICA basis and therefore shows estimated component counts in the intervals [1,2), [2,3), [3,4), [4,8), [8,16), [16,32), and [32,infinity). Panels: gpt2 (GPT-2 small), gemma2 (Gemma 2 2B), qwen9b (Qwen 3.5 9B Base).
