Categories Companies Research

Loading the SOTA2 catalog…

Explore

Categories Companies Research

Contribute

Join SOTA2 Contact

SOTA2

© 2026 SOTA2

Build in Bellevue, USA

WizWand is now SOTA2 Research

Papers Benchmarks Datasets Tasks

Sparse Autoencoder Evaluation research benchmarks · SOTA2 Research

Sparse Autoencoder Evaluation

Benchmarks

Dataset NameSOTA MethodSort

Dataset NameSOTA methodMetricTrendResultsLast Updated

Gemma-2-2B activations

320L0 Count

20

May 8, 2026

DINOv2-B activations

87.4Variance Explained

6

Apr 30, 2026

DINOv2 S activations

94.2Variance Explained

6

Apr 30, 2026

Qwen3-8B layer 20

0.989KL Score

4

Jun 30, 2026

SAEBench intrinsic evaluation Gemma 2-2B

Ensembling (NB)

0.763RAVEL Score

4

Jun 15, 2026

SAEBench Pythia-160M (test)

Ensembling (Boosting)

0.502RAVEL Score

4

Jun 15, 2026

SAEBench Pythia-70M

Ensembling (Boosting)

0.308RAVEL Score

4

Jun 15, 2026

Pythia-160M activations SAEBench downstream performance

3.906CE Loss

4

Jun 1, 2026

Most-probable construction SAE

67.3Top-1 Sparse Probing Accuracy

2

Jun 11, 2026

Popular datasets

SAEBench Qwen Dino Gemma DINOv2

Follow for updates

Join SOTA2 Contact