Can we remove EM by ablating neurons?
On the EM-positive sparse mask (Qwen3.5-9B, midsp λ=3e-5 ep10; baseline EM 1.0%, mean-align 77.2). Revert = set a neuron's mask edit back to identity (M→1). Higher mean-align = less EM.
Revert each neuron set (K=60)
| revert set | EM% | mean-align | al<30 |
|---|
| baseline(full mask) | 1.0 | 77 | 18 |
| interp | 1.0 | 85 | 6 |
| cpm_pm | 0.9 | 86 | 2 |
| absA | 0.5 | 87 | 3 |
| maskmag | 0.5 | 86 | 4 |
| random | 0.5 | 88 | 4 |
| allchanged | 0.0 | 97 | 0 |
No selective removal: interpreted (amplified⊕flipped), C+/C−, and |grad×act| reverts do NOT beat the random control (all move mean-align 77→85–88). Only reverting all changed neurons (undoing the whole edit) removes EM (mean-align 96.6, EM 0).
K-sweep — does attribution beat random at larger K?
Revert the top-K changed neurons ranked by each method, K=20→5000. Baseline 77.2, full-revert 96.6.
| rank by ↓ / K → | 20 | 60 | 150 | 500 | 1500 | 5000 |
|---|
| C+/C− | 86.7 | 86.4 | 85.0 | 86.9 | 86.1 | 84.7 |
| |grad×act| | 85.2 | 86.9 | 86.2 | 86.0 | 86.8 | 83.7 |
| |M−1| (edit size) | 86.9 | 85.7 | 85.5 | 86.2 | 88.8 | 90.8 |
| random | 85.0 | 88.0 | 86.0 | 83.1 | 85.5 | 85.6 |
Attribution on the narrow training signal never beats random. Ranking neurons by C+/C− and |grad×act| of the insecure-code-completion target stays flat ~84–87 up to K=5000. Only |M−1| (raw edit magnitude) separates (90.8 @K=5000) — a bulk effect (undo the biggest edits), not an attribution signal.
Attribute the RIGHT target: MCQ, on the items that actually show EM
The free-form (generative) attribution and its ablation results were removed 2026-07-31. EM attribution is MCQ-only now, and conditioned on the probe items where the model actually picks the misaligned option. Numbers for this section are regenerated by em_mcq_empos.sh.
Oracle skyline search — does a SMALL removable set exist?
Greedy eval-set search directly on the MCQ objective (not attribution): screen the biggest-edit neurons, revert best-first up to k=200 — the strongest possible localization.
| k reverted | 0 | 25 | 50 | 100 | 150 | 200 |
|---|
| MCQ divergence | -2.5774 | -2.7902 | -2.8353 | -2.9055 | -3.0222 | -3.0539 |
No — even the oracle can't localize a small set. Greedy search moves MCQ divergence only −2.58→−3.10 at k=189 (~40% toward base −4.4), and on the FAITHFUL judge its sets TIE random: mean-align 86.6/86.7/85.8 at k=50/100/200 (vs random ~86–88). So no ≤200-neuron removable set exists even when searching the eval directly — decisively confirming EM is distributed; removable only in bulk (thousands of neurons), never localized to a small interpretable set.