SAE-based feature-level interventions to suppress PII leakage in LLMs are primarily evaluated by log-probability reduction. We show that this criterion cannot distinguish feature specific suppression from a magnitude artifact: a direction-agnostic effect of sufficiently large activation perturbations. Across 36 (model, layer, PII type) cases, we compare selected SAE feature interventions with coherencematched random direction controls using logprob and rank-aware metrics. While 17 cases (47%) appear successful under logprob-only evaluation, only 1 case (2.8%) passes our multimetric framework. The rest reveal two failure modes: random directions outperform selected features, and feature selection collapse across PII types. These results indicate that existing evaluations cannot verify whether interventions target causal pathways of PII leakage. A 9,000 run causal mediation feasibility study corroborates this: the verifier separates candidates reliably, but correlation-selected candidates rarely pass it. We therefore propose a new research mission for NLP: causally grounded PII suppression through circuit-level localization and causal mediation analysis.
Figures & tables
Figure 1: Main diagnostic results across 36 (model, layer, PII type) cells. (a) Number of cells passing single-metric vs. multi-metric criteria. (b, c) Per-template Δlp for selected feature vs. coherence-matched random control. (d) ∣Δlp∣ per PII type in Llama L8, where f=16363 is selected for all four types.
Model
PII
TE
NIEf
NIErand
Ratio
Indist.
Gemma
email
+2.580
2.504
2.574
0.973
100.0
id
−0.413
−0.420
−0.399
1.053
100.0
name
+0.609
0.408
0.649
0.628
73.3
Llama
email
+3.560
3.558
3.559
1.000
100.0
id
−0.233
−0.241
−0.237
1.016
100.0
name
−0.283
−0.290
−0.283
1.023
100.0
Table 1: Causal mediation feasibility for (model, PII type) combinations with ∣TE∣≥0.05 (150 of 180 candidate features). NIE f and NIE rand are means over 15 candidate features; Ratio is their ratio, and Indist. is the percentage of candidates whose individual ratio lies in (0.5,2.0) . Bold marks the only cell in which some candidates are distinguishable from random; it is also the only non-email cell with a positive memorization signal. Full protocol in Appendix A.14 .
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Depth
dmodel
SAE layers
Gemma-2-9B
42
3584
9, 20, 31
Llama-3.1-8B-Instruct
32
4096
8, 16, 24
Qwen3-8B
36
4096
9, 18, 27
Appendix
Table 2: Model and layer configuration. SAE layers are selected at approximately 25%, 50%, 75% depth.
Split
Pair indices
Purpose
Discovery
0–9
Feature discovery
Diagnostic
0–19
Reconstruction diagnostics
Validation
300–349
Main multi-metric evaluation
SAE training
separate 50k corpus
Task-SAE training
Appendix
Table 3: Data split used for feature discovery, diagnostics, and final evaluation. The SAE training corpus is entity-level disjoint from the 600 evaluation pairs.
Hyperparameter
Value
Expansion factor
×8
Sparsity k
32
Auxiliary loss weight
0.25
Dead feature threshold
1000 inactive steps
Training steps
80,000
Batch size
4,096 tokens
Appendix
Table 4: Task-specific SAE training hyperparameters.
Model
Layer
Rel. err. all
Rel. err. PII
Cosine PII
Gemma-2-9B
9
5.11%
2.78%
0.9996
Gemma-2-9B
20
10.28%
6.38%
0.9980
Gemma-2-9B
31
9.74%
8.82%
0.9961
Llama-3.1-8B-Instruct
8
5.20%
1.99%
0.9998
Llama-3.1-8B-Instruct
16
6.77%
13.53%
0.9907
Llama-3.1-8B-Instruct
24
10.45%
39.20%
0.9192
Appendix
Table 5: SAE reconstruction quality across nine (model, layer) cells. Relative error is ∥h−h^∥/∥h∥ ; "all" averages over all token positions, "PII" over PII token positions only. Cosine similarity is computed at PII token positions. Deep layers (Llama L24, Qwen L27) show higher PII-position reconstruction error.
ID
Example format for name PII
Context
T0
Patient Name: {name}
all 4 PII fields
T1
Patient Name: {name}
target PII only
T2
Name: {name}
target PII only
T3
Full name: {name}
target PII only
T4
The patient’s name is {name}
target PII only
T5
환자 이름: {name}
target PII only
Appendix
Table 6: Prompt templates used for cross-template evaluation. T0 retains the original training prompt that contains all four PII fields (name, patient ID, phone, email) together; T1–T8 isolate the target PII field only, with varying label formats. T8 is abbreviated (full text: “The patient, {name}, presented with chest pain.”). The same template family is adapted to patient ID, phone number, and email PII.
Criterion
Meaning
Δlp≤−0.3
logprob suppression
Δbr≥+1.0
rank-cost improvement
Δlp>+0.1
logprob anti-suppression
Δbr<−0.3
rank-cost degradation
Sellp,Selbr≥5
strong feature-specificity
Appendix
Table 7: Five criteria used for category labels.
Category
Definition
clean-strong
Δlp≤−0.3 , Δbr≥+1.0 , Sellp,Selbr≥5
clean-weak
Δlp≤−0.3 , Δbr≥+1.0 , Sel<5 on ≥ 1 metric
anti
Δlp>+0.1 or Δbr<−0.3
lp-only
Δlp≤−0.3 but Δbr<+1.0
br-only
Δbr≥+1.0 but Δlp>−0.3
inactive
zt[f]≈0
Appendix
Table 8: Category definitions used for post-hoc labeling. Each cell is assigned exactly one category. Raw Δ and Sel values are reported alongside category labels.
Figure 2: Selected feature intervention effect ( ∣Δlptarget∣ , y-axis) versus coherence-matched random direction effect ( ∣Δlprandom∣ , x-axis), per (model, layer, PII type) configuration; both axes log scale. Large markers indicate cell mean over 9 templates × 50 validation pairs (450 evaluation contexts); small dots indicate per-template means within a cell. Gray dashed: Sellp=1 (target = random). Orange dashed line: Sellp=5 (clean-strong selectivity threshold). Diamond ( ⧫ ): cell satisfying all clean-strong criteria; hollow markers: inactive configurations ( zt[f]≈0 ). Of 36 configurations, only Llama L16/name qualifies as clean-strong. Top-left tag in each subplot encodes category labels for [name, id, phone, email] (N=noise, W=clean-weak, A=anti, L=lp-only, B=br-only, ★ =clean-strong). The clean-strong label ( ⧫ ) reflects the full multi-metric criterion including rank-aware Δbr and signed-metric checks.
Model
Layer/PII
ΔlpT
ΔlpR
Sellp
ΔbrT
ΔbrR
Selbr
Category
Gemma
9/name
−0.005
+0.009
0.52
+0.053
−0.079
0.68
noise
Gemma
9/id
+0.021
−0.123
0.17
+0.111
+0.310
0.36
noise
Gemma
9/phone
−0.235
−0.060
3.95
+0.144
−0.040
3.56
noise
Gemma
9/email
+0.002
+0.003
0.88
+0.007
+0.004
1.54
noise
Gemma
20/name
−2.805
−7.524
0.37
+4.806
+12.906
0.37
clean-weak
Gemma
20/id
−0.475
−3.460
0.14
+3.166
+18.171
0.17
clean-weak
Appendix
Table 9: Full per-cell results over 36 configurations. T = target (selected feature); R = coherence-matched random. All values are means over 50 validation pairs × 9 templates. Inactive configurations have zt[f]≈0 across all templates, so neither effect nor selectivity is well-defined.
Template
ΔlpT
ΔlpR
Sellp
Llama L16/name (clean-strong; positive control)
T0 (in-distribution)
−1.153
−0.0003
3791
T1
−1.155
−0.0004
3261
T2
−1.337
−0.0004
3266
T3
−1.346
−0.0004
3549
T4
−1.078
−0.0007
1449
Appendix
Table 10: Per-template results for cells discussed in the main text. T0 retains the original training format containing all four PII fields; T1–T8 isolate a single PII field with varying label formats. Llama L16/name maintains Sellp>400 across all 9 templates including OOD Korean and JSON formats. Gemma L20/name shows Sellp<1 across all 9 templates (Failure mode A).
Model
Layer
name
id
phone
email
Full feature-selection collapse
Gemma
20
20385
20385
20385
20385
Llama
8
16363
16363
16363
16363
Qwen
9
30639
30639
30639
30639
Partial feature sharing
Gemma
9
5559
5559
4648
11893
Appendix
Table 11: Selected feature ID per (model, layer, PII type). Group definitions: Full collapse =single selected feature across all 4 PII types; Partial sharing =some PII types share a selected feature; No sharing =distinct selected feature for all 4 PII types. SAE dictionary size: dsae=28,672 for Gemma, 32,768 for Llama and Qwen. Bolded : feature 32619 (Llama L16/name) corresponds to the only clean-strong configuration in the entire grid.
Configuration
Selected Δlp
Decoder control Δlp
Stronger
Llama L16/name
−1.06
−0.007±0.014
0/20
Llama L8/name
−4.57
−6.60±3.48
14/20
Qwen L18/name
−1.38
−1.04±0.30
2/20
Gemma L20/name
−2.80
−9.10±6.38
20/20
Appendix
Table 12: Random SAE decoder control on four name configurations spanning the main diagnostic categories. Decoder control values are mean ± s.d. over 20 sampled directions. “Stronger” is the number of control directions yielding a more negative Δlp (stronger suppression) than the selected direction.
Collapse cell
name
id
phone
email
Gemma L20 (f=20385)
−2.80
−0.47
−0.45
−0.31
Llama L8 (f=16363)
−4.57
−1.96
−1.95
−1.97
Qwen L9 (f=30639)
−1.92
+0.16
−0.48
+1.11
Appendix
Table 13: Δlptarget per PII type in the three full-collapse cells. Negative values indicate suppression; positive values indicate anti-suppression. Despite shared feature selection and matched activation magnitudes, intervention effects are markedly type-asymmetric: Gemma L20 favors name (6.8 × average of others), Llama L8 favors name (2.3 × ), and Qwen L9 produces anti-suppression for id and email.
Threshold
Single-metric
Dual-metric
≥2
3 (8.3%)
3 (8.3%)
≥3
2 (5.6%)
2 (5.6%)
≥5
2 (5.6%)
1 (2.8%)
≥10
2 (5.6%)
1 (2.8%)
Appendix
Table 14: Selectivity threshold sensitivity. Single-metric: Sellp only; Dual-metric: both Sellp and Selbr above threshold. Default magnitude gates ( Δlp≤−0.3 , Δbr≥+1.0 ) and anti-suppression filter retained. Single-metric at ≥5 retains Llama L8/phone ( Sellp=11.2 , Selbr=3.4 ), which dual-metric filters out.
∣ΔlpT∣ gate
ΔbrT gate
Cells / 36
≥0.0
≥0.0
2
≥0.0
≥0.5
1
≥0.3
≥1.0 (default)
1
≥0.5
≥1.0
1
Appendix
Table 15: Magnitude gate sensitivity at fixed Sel≥5 . Even the loosest gates yield only 2/36 (5.6%) clean-strong cells.
Figure 3: α sensitivity on three representative cells. Yellow band: main-paper setting ( α=−2 ). (a) Llama L16/name (clean-strong): selected-feature dominance is preserved across all tested α . (b) Gemma L20/name (Failure mode A): random-direction dominance is preserved across all α , including the over-perturbation regime at α=−5 . (c) Qwen L18/name (clean-weak): sellp exceeds 1 at α=−1,−2 but crosses below 1 from α=−3 onward, indicating that the weak selectivity observed in the main experiment is itself α -dependent.
sellp as a function of α (T1+T5)
Cell
−1
−2
−3
−4
−5
Llama L16/name
—
1507
2486
320
39
Gemma L20/name
1.31
0.41
0.23
0.18
0.50
Qwen L18/name
1.31
1.54
0.83
0.53
0.52
Appendix
Table 16: Selectivity ratio sellp across α for three representative cells: Llama L16/name (clean-strong), Gemma L20/name (Failure A), Qwen L18/name (clean-weak). Bold marks the α at which selectivity flips below 1. Llama L16/name α=−1 omitted ( ∣Δlprandom∣ at measurement floor).
Configuration
PubMedQA
MedQA
MMLU
Llama L16/name (clean-strong)
31.0→32.0
45.5→46.5
61.0→61.5
Qwen L18/name (clean-weak)
28.5→16.5
57.0→23.5
71.5→48.5
Gemma L20/name (Failure A)
47.5→33.5
60.5→34.5
67.5→45.0
Appendix
Table 17: Downstream accuracy (%) before → after intervention. Llama L16/name, the only clean-strong cell, shows no degradation on any benchmark, whereas both non-clean-strong cells degrade consistently across all three.
PII type
Gemma-2-9B
Llama-3.1-8B
Qwen3-8B
name
+0.609
−0.283
−0.604
id
−0.413
−0.233
−0.025
phone
−0.038
+0.067
−0.065
email
+2.580
+3.560
+2.466
Appendix
Table 18: Total effect (TE) by PII type and model, averaged over 50 counterfactual pairs per (model, PII type) cell. TE is defined independently of any feature intervention or aggregation choice.
Privacy leakage in language models concerns how models expose training data, infer sensitive traits, or transmit private context through outputs and internal channels.