Sampled-SAE pre-selects a candidate pool of features using batch-level norms or entropy before batch top-K, creating a tunable family that trades reconstruction fidelity for improved probing and reduced absorption on Pythia-160M.
High frequency latents are features, not bugs
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Distribution-Aware Feature Selection for SAEs
Sampled-SAE pre-selects a candidate pool of features using batch-level norms or entropy before batch top-K, creating a tunable family that trades reconstruction fidelity for improved probing and reduced absorption on Pythia-160M.