Last released Jul 17, 2026
PrefScope: an interpretability framework for diagnosing post-training data and model behavior with sparse autoencoders.
Supported by