An audit of SAEBench reveals that Targeted Probe Perturbation and Spurious Correlation Removal metrics fail reliability tests and should not be used to evaluate sparse autoencoders.
Ravel: Evaluat- ing interpretability methods on disentangling language model representations
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Are Sparse Autoencoder Benchmarks Reliable?
An audit of SAEBench reveals that Targeted Probe Perturbation and Spurious Correlation Removal metrics fail reliability tests and should not be used to evaluate sparse autoencoders.