sparse autoencodersfeature attributionpublished Sep 30, 2026 · arXiv:2609.35890
Representational Simplicity and Circuit Size Dissociate in a Threshold-Dependent Way: A Controlled Test via Adversarial Training
mechanistic interpretabilitydesigned measurementpublished Aug 21, 2026 · arXiv:2608.19338
Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability
mechanistic interpretabilitycircuit discoverypublished Aug 17, 2026 · arXiv:2608.13754
Explanation Multiplicity: Circuit-Level Interpretability Evidence Does Not Survive Defensible Analytic Variation