Reducing Logprob Without Causal Verification: Diagnosing SAE-Based PII Suppression in LLMs
| Source: arXiv AI
Tags: Sparse Autoencoders, PII Leakage, Mechanistic Interpretability, Causal Mediation Analysis, AACL-IJCNLP 2026
A paper finds that SAE-based PII suppression in LLMs looks successful in 47% of cases by log-probability but passes a stricter multi-metric test in just 1 of 36, often doing no better than random directions.
Details
Sparse autoencoder feature interventions to stop LLMs leaking personal data are usually judged by how much they reduce log-probability of the PII. The authors show this cannot separate feature-specific suppression from a generic effect of large activation perturbations. Across 36 model, layer and PII-type cases, they compare chosen SAE features against coherence-matched random directions. 17 cases (47%) look successful on logprob alone, but only 1 (2.8%) passes their multi-metric framework. Failures include random directions outperforming selected features and the same features being picked across PII types. A 9,000-run causal mediation study supports this. The authors call for circuit-level localization and causal mediation analysis for PII suppression.