Web

Beyond Reconstruction: Verifying Model Explanations with RECAP

What Changed For years, the field of mechanistic interpretability has relied heavily on natural-language autoencoders to translate hidden model acti…