MIB, a multi-task benchmark for mechanistic interpretability, reports that attribution and mask optimization methods excel at circuit localization, and DAS outperforms sparse autoencoder features at causal variable localization.
Title resolution pending
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
-
MIB: A Mechanistic Interpretability Benchmark
MIB, a multi-task benchmark for mechanistic interpretability, reports that attribution and mask optimization methods excel at circuit localization, and DAS outperforms sparse autoencoder features at causal variable localization.