Reducing Logprob Without Causal Verification: Diagnosing SAE-Based PII Suppression in LLMs
This work proposes a new research mission for NLP: causally grounded PII suppression through circuit-level localization and causal mediation analysis, indicating that existing evaluations cannot verify whether interventions target causal pathways of PII leakage.
Chaerin Lee, Wonho Lee, Sohee Park et al.
· 0 citations