[HUGGINGFACE]score: 0.48

Stable SAE Features Carry Most Signal; Unstable Features Are Low-Frequency Noise

June 9, 2026

Across seeds, models, layers, and dictionary sizes, stable sparse autoencoder features account for most reconstruction and prediction-relevant signal, while unstable features are dominated by low-frequency surface-form triggers. The study introduces a scalable per-feature stability score to separate the two.

HOW THIS AFFECTS YOU

●

researcherThe per-feature stability metric gives interpretability researchers a practical filter for which SAE features are worth analyzing, reducing noise in mechanistic studies.

●

policyWorth watching because SAE-based circuit analysis used in safety research may be systematically biased toward unstable, low-signal features without this filtering step.

read original ↗huggingface.co

← back to feed