Concept-level explanation methods land in a dedicated AI alignment track
AAAI 2026's Special Track on AI Alignment features work such as PCMNet, which supplies concept-level explanations aimed at transparency, controllability and trustworthiness — an attempt to make model internals legible at the level humans actually reason about.
Feature-level interpretability produces explanations in the model's vocabulary. Concept-level methods try to produce them in ours, and the gap between those two is why interpretability results are often more impressive to researchers than useful to operators.
The pitch is controllability rather than description. If a system's behaviour can be attributed to concepts a human can name, those concepts become a surface you can intervene on — which is a stronger claim than explanation and a much more useful one for anyone who has to sign off on a deployment.
The standing objection is that a concept legible to a human may not be a natural unit inside the model. Imposing a vocabulary produces explanations that are satisfying and possibly wrong, which is the worst combination available. Whether the concepts are recovered or projected is the question every method in this class has to answer.
The venue is part of the news. A dedicated alignment track at a mainstream AI conference means this work is being reviewed by the same process as everything else, rather than circulating in a parallel community — and 2026's papers are noticeably more empirical, with controlled experiments and formal user studies where there used to be argument.
AAAI 2026 — Special Track on AI Alignment → · arXiv — Open Problems in Mechanistic Interpretability →