rome-restoration-run-length-10-layers
IN premise — summaries/2026/08/24/meng-2022-rome-sR-references.md
Created 2026-08-25T02:58:17+00:00
Decisive causal effects in Causal Tracing require restoring a run of 10 consecutive layer activations (interval [l*−4, ..., l*+5]); single-vector restorations have negligible effects.
Summary
Causal Tracing only produces meaningful results when you restore a window of ten consecutive layer activations around the critical layer; poking a single layer in isolation gives almost no signal. This means the computation is genuinely spread across a span of layers, so any method that tries to pin causal responsibility on one layer alone will miss what is actually happening.