DECODE adds human oversight to LLM-based interpretability pipelines
The arXiv paper introduces a visual analytics system that diagnoses where automated interpretation of large language models goes wrong, using consensus among explanations as a reliability signal.
In plain words
Researchers increasingly use large language models to both produce and judge explanations of how other models work internally, with little human checking. DECODE is a visual tool that brings people back into that loop. It breaks the diagnostic work into stages and flags which instances deserve manual review.
What was announced
The authors argue that when the same class of model generates and evaluates explanations, reliability suffers and failures are hard to trace, since it is unclear whether a defect lies in the extracted representations or in the LLM doing the interpreting. DECODE addresses this with a three-stage workflow that isolates failure sources, from structural defects to misaligned explanations, and borrows the idea of inter-coder consensus from qualitative research: agreement among independently generated explanations and their metrics serves as a proxy for reliability.
To keep human effort manageable, a three-phase active seeding protocol recommends high-priority instances for manual review. In a case study covering 16,384 sparse autoencoder features, the authors report that the system identified and attributed interpretability failures at scale. The evidence is limited to the paper's abstract, so the case study's methods and results cannot be independently verified from this source.
What it does not show
- The source is only the paper's abstract; full methods, datasets and results are not available for verification.
- The claimed effectiveness rests on the authors' own case study of 16,384 sparse autoencoder features, with no independent confirmation.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.