memit-causal-tracing-3sigma-noise
IN premise — summaries/2026/08/24/meng-2022-memit-sR-references-chunk-1.md
Created 2026-08-25T02:58:14+00:00
Causal tracing for MEMIT layer selection injects Gaussian noise at 3σ (empirical embedding variance) into subject-token hidden states, then restores individual layer outputs to measure causal contribution.
Summary
To decide which layers of the network actually carry a given memory, MEMIT deliberately scrambles the subject token's internal representations with a carefully scaled amount of noise, then fixes one layer at a time and checks which restoration brings the output back on track. This tells the system exactly where to aim its edit, so changes are made only to the layers that genuinely matter for that specific fact rather than being spread indiscriminately across the whole model.