INT8 · NF4 · damage-rank survival
Base-model quantization noise can swamp edit damage-rank.
Locate-then-edit efficacy and the rank-order of collateral damage are re-scored after INT8 and NF4 quantization. Full-model NF4 injects diffuse base noise; edited-layer-only NF4 and INT8 preserve much more of the unquantized rank.
The object
Noise-to-signal is the ratio of quantization-induced logit jitter to the edit’s own damage signal. Rank-survival is the Spearman between unquantized and quantized damage ranks.
The curve asks whether families with higher noise-to-signal lose damage-rank first. That is a measurement about base-model noise, not a new editor.
What is measured
- Local curve, 9/9 cells: Spearman(NSR, NF4 rank-survival) = −0.900
- Family means (rank-survival): Gemma-2-2B 0.934 · Llama-3.2-1B 0.904 · Llama-3.2-3B 0.680 · Qwen-1.5B 0.276 · Qwen-3B 0.205 · Phi-3.5 0.128
- Qwen NSR rises with scale (monotone); Llama is only partially monotone
- Edited-layer-only NF4 and INT8 keep rank structure that full-model NF4 erases
Protocol
- Install a ROME (or listed editor) edit on the unquantized model; dump damage ranks.
- Re-load the same weights under INT8 / NF4 (full model or edited layer only).
- Re-score efficacy and damage ranks; plot survival against noise-to-signal.
Measurements
Swamping
When base-model noise is large relative to the edit signal, pairwise damage order collapses. The −0.900 curve is that statement.
Family split
Gemma and Llama-1B stay high; Qwen and Phi sit much lower. The split is family-structured, not a single scalar.
Open cells
An 8B anchor and a channel-scale × NF4 gap axis are not closed on this page. Do not infer them from other 8B geometry cells.