ResearchBenchmarksSparse Autoencoder Evaluation on Pythia-160M activations SAEBench downstream performanceSparse Autoencoder Evaluation on Pythia-160M activations SAEBench downstream performanceFollow3.906CE LossiSAE-ME3.902083.928543.9553.98146May 29, 2026Evaluation resultsMethodMethodLinksCE LossSparse Probing AccuracySCRTPPiSAE-MEBackbone=Pythia-160M,...Backbone=Pythia-160M, Encoding scheme=multi-step2026.053.90690.829.34.4TopKBackbone=Pythia-160MBackbone=Pythia-160M2026.053.97490.938.14.3iSAEBackbone=Pythia-160M,...Backbone=Pythia-160M, Regularization weight=10^-12026.054.00390.734.84.3AbsTopKBackbone=Pythia-160MBackbone=Pythia-160M2026.054.00490.6333.1