AI organizes its skills along an orthogonal basis that bears no resemblance to human categories.
April 23, 2026
Original Paper
Characterizing Model-Native Skills
arXiv · 2604.17614
The Takeaway
We usually describe AI capabilities using human terms like coding or translation. This research shows that the model internal activations follow a completely different set of coordinates. Recovering this native basis allows for much more precise steering of the model behavior. We can activate or suppress specific skills by targeting the model own internal map. Using human taxonomies to control AI is like using a map of New York to navigate Mars. This discovery lets us interact with the model on its own terms for the first time.
From the abstract
Skills are a natural unit for describing what a language model can do and how its behavior can be changed. However, existing characterizations rely on human-written taxonomies, textual descriptions, or manual profiling pipelines--all external hypotheses about what matters that need not align with the model's internal representations. We argue that when the goal is to intervene on model behavior, skill characterization should be *model-native*: grounded in the model's own representations rather t