Megadose The AI curve, compressed.

DECODE: Diagnosing Interpretability Failures in Automated Interpretation Process

· ArXiv · AI/CL/LG ·
DECODE helps human reviewers trace whether interpretation failures come from extracted representations or the LLM explaining them.

The visual analytics system divides diagnosis into three stages, covering structural defects through mismatches in explanations. Without ground truth, it uses agreement among independently generated explanations and their evaluation metrics as a proxy for reliability. A three-phase protocol prioritizes instances for manual review. The authors report that a case study covering 16,384 sparse autoencoder features demonstrated efficient identification and attribution of failures at scale.

ArXiv · AI/CL/LG's note

score 4

Categories: Research