@inproceedings{lee-etal-2026-reducing,
title = "Reducing Logprob Without Causal Verification: Diagnosing {SAE}-Based {PII} Suppression in {LLM}s",
author = "Lee, Chaerin and
Lee, Wonho and
Park, Sohee and
Choi, Daeseon",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.182/",
pages = "3636--3654",
ISBN = "979-8-89176-444-6",
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 coherence-matched 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 multi-metric 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."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lee-etal-2026-reducing">
<titleInfo>
<title>Reducing Logprob Without Causal Verification: Diagnosing SAE-Based PII Suppression in LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chaerin</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wonho</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sohee</namePart>
<namePart type="family">Park</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Daeseon</namePart>
<namePart type="family">Choi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<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 coherence-matched 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 multi-metric 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.</abstract>
<identifier type="citekey">lee-etal-2026-reducing</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.182/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>3636</start>
<end>3654</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Reducing Logprob Without Causal Verification: Diagnosing SAE-Based PII Suppression in LLMs
%A Lee, Chaerin
%A Lee, Wonho
%A Park, Sohee
%A Choi, Daeseon
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F lee-etal-2026-reducing
%X 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 coherence-matched 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 multi-metric 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.
%U https://aclanthology.org/2026.findings-aacl.182/
%P 3636-3654
Markdown (Informal)
[Reducing Logprob Without Causal Verification: Diagnosing SAE-Based PII Suppression in LLMs](https://aclanthology.org/2026.findings-aacl.182/) (Lee et al., Findings 2026)
ACL