Interpretability beyond human concepts
Published on September 22, 2026, the paper investigates whether language models use internal representations that do not map adequately to familiar human concepts.
Source: Xeno-Interpretability e Representações Nativas de Modelos (arxiv.org). Text prepared with AI from this source.
What happened and what to do
Published on September 22, 2026, the paper examines whether language models use internal distinctions for which humans have no adequate concepts. It calls these structures “xeno-representations” and their analysis “xeno-interpretability”; its focus is the possible limit of interpreting model representations through familiar human categories.
For a company, this question can become an evaluation process: define relevant tasks and behaviors, record model outputs, and check whether explanations based on familiar concepts account for what was observed. A team can organize this data collection, establish review criteria, and track results in dashboards without assuming that a familiar explanation fully describes a model’s internal workings.
How the consultancy can help
Wendelmaques can diagnose an interpretability opportunity or exposure, define an evaluation scope, and implement the necessary data collection, review criteria, and dashboards. Operations can also be structured around the company’s process.
Next step
Send a short description of your case to receive a scoped proposal for evaluating model representations and organizing results tracking.
Let's assess your project
I reply with scope and a proposal.