A good interpretation
tells you what to do next.

Interpretability researchers have spent years developing techniques to look inside language models and neural networks. Which circuits handle factual recall? Which attention heads track syntactic relationships? Which neurons activate on specific concepts? The work is technically careful, often surprising, and has produced a genuine body of knowledge about how models process information.

But there is a gap between having an interpretation and being able to do something with it. If a researcher shows that a set of neurons activates on gender-stereotyped associations, the useful follow-up question is: what should I change, and does changing it actually fix the problem? An interpretation that cannot answer those questions is interesting. It is not yet a tool.

The problem this position paper names is that the field has largely been evaluating interpretability research on a different standard: whether an explanation is internally coherent, predictive of some behavioral signal, or elegant in its account of what the model does. That is a real scientific standard. But it is not sufficient to make a finding actionable, and the authors argue that the field has built most of its reward structure around the first standard without enough pressure on the second.

The rubric creates four quadrants. Low concreteness and low validation is exploratory research: valuable for building understanding but not a basis for production decisions. High concreteness and low validation is a proposed intervention awaiting a test. High concreteness and high validation is genuinely actionable: a finding that tells you what to change and shows that the change works.

The paper also identifies five domains where interpretability that clears both criteria provides unique leverage beyond what behavioral testing can deliver. Behavioral testing tells you how a model performs on a distribution of tasks. It cannot easily tell you why a specific failure pattern is happening internally or how to fix it at the root. When interpretability is both concrete and validated, it can support root-cause interventions that behavioral benchmarking cannot.

This is a position paper, which means the evidence comes from literature review and argument rather than from a new controlled experiment. The authors read existing interpretability work through the lens of their two-dimension rubric and report what they find.

The pattern, they argue, is that the field has produced more work that is concrete than work that is validated. Researchers have become skilled at identifying specific components, directions, or circuits inside models that correlate with specific behaviors. That is the concreteness criterion. What has received less consistent attention is whether proposed interventions on those components actually produce the expected change in real-world use, without degrading other model capabilities. That is the validation criterion.

Some work is concrete but unvalidated: the proposed intervention is specific, but its downstream effects have not been measured across a realistic use-case distribution. Other work is validated in a narrow sense (it predicts some behavioral signal in a controlled setting) but too abstract to propose a specific real-world action. The paper argues the field needs both, and that grading work on only the first dimension has created a gap between how interpretability is evaluated in research and how it would need to perform to be trusted in production.