arXiv:2611.04782v1 Announce Type: new
Abstract: 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.
arxiv
Logprob Reduction Fails to Verify Causal PII Suppression in LLMs
This paper shows that SAE-based feature interventions for suppressing PII leakage in LLMs, when evaluated only by log-probability reduction, often produce false positives. Across 36 cases, 47% appeared successful under logprob-only evaluation, but only 2.8% passed a multimetric framework with random direction controls. The authors identify two failure modes and propose causally grounded PII suppression via circuit-level localization and causal mediation analysis.
Provides concrete evidence that a common evaluation metric is unreliable)Skip and proposes a more rigorous methodology, directly impacting AI safety research.
Comments & Questions (0)
No comments yet
Be the first to comment!