Interpretability Needs a Machine Neurology
Researchers are now examining the internal activity of language models to explain behaviors such as false statements, flattery, and harmful advice. We propose a way to report every claim about the inside of a model like a clinical diagnosis, along with the examination performed, what it rules out, and the explanations it leaves open. Published results have shown that each common examination supports less than its usual reading. A classifier can read a concept that the model never uses, an intervention can alter the answer through a pathway that stays idle on ordinary inputs, and removing a component can leave the behavior intact because other components take its place. We collect these results into one table listing what twelve examinations can and cannot establish, and we organize the existing literature along the steps of clinical work, from symptoms and examinations to localization, differential diagnosis, cause and treatment. We audit 60 papers from ICML 2026 and find that all of them state how the examination was set up, while 50% indicate an alternative explanation of their main findings, 40% test one such explanation, and 8% perform the examination on a case whose answer was known. We call this clinical approach machine neurology and propose a six-item report that authors can write and reviewers can request.
Authors
- Hanyao Zeng
- Victor Sheng
- Xin Zhang (ORCID: https://orcid.org/0009-0009-8935-0720)
Institutions
- Texas Tech University (US)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24422422
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- preprint