Study finds structural shortcuts in automated interpretability evaluation data
A new paper examines how automated interpretability scores SAE feature explanations by predicting token activations, and identifies structural shortcuts in the evaluation data that allow scoring without reading explanations.
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-10-30 04:00 · arXiv cs.CL
Scoring Explanations Without Reading Them: Structural Shortcuts in Automated-Interpretability Evaluation Data