fv-gptneoxx-middle-layer-clustering
IN premise — summaries/2026/08/24/todd-2023-function-vectors-s2-function-vector-causal-effects-cannot-be-recovered-from-the--chunk-3.md
Created 2026-08-25T02:58:41+00:00
High-AIE heads cluster in middle layers across GPT-J, Llama 2, and GPT-NeoX, with GPT-NeoX being an exception that clusters in earlier-middle layers (10–20).
Summary
Across three major transformer families, the attention heads that matter most (by this influence metric) consistently land in the middle of the network rather than at the input or output ends, with GPT-NeoX sitting slightly earlier in that middle band. This matters because it points to a shared architectural principle: the network's critical interpretive work happens in a predictable middle zone, so anyone trying to inspect, edit, or control model behavior should focus there first.