SAE Feature Interventions Not Uniformly Localized for Safety Control, New Evaluation Shows

Jul 14, 2026

A new study introduces a matched coherence-gated evaluation protocol for sparse autoencoder (SAE) features in safety interventions. Testing on Gemma-2-9B-it, the authors find that SAE feature ablation has a narrow useful regime, with higher-rank features causing coherence collapse rather than localized control. The results suggest SAE-based safety interventions should be evaluated as regime-dependent mechanisms.

Why it matters: This challenges the assumption that SAE features are uniformly localized control handles, which is critical for developing reliable AI safety interventions.

Full story at: arXiv AI/ML ↗