Definition
Plain language
The full set of problems a model can eventually crack given enough tries, no matter how many attempts you allow it.
As stated in the literature
In pass-at-k-style analysis, the limit of what a model can solve as the number of independent attempts grows without bound; capability expansion means one model's boundary contains problems another's does not at any sampling budget, distinct from mere sampling efficiency.
Also called: capability boundaries
Why it matters: It matters because it distinguishes a model that is truly more capable from one that just needs more attempts to reach the same ceiling.
For example, one model might eventually solve a tricky puzzle if allowed thousands of tries, while another simply never can no matter how many attempts it gets.
Heard on the show
“That's the capability boundary.”Episode 011 — When RL Actually Teaches Agents Something New, And When It Doesn't