capability

Definition

What a model can actually do, usually operationalised as scores on benchmarks that probe a specific skill. Scaling papers ask whether new capabilities appear as models grow; evaluation papers argue about whether a benchmark measures the capability it names. In safety discussions the term is kept deliberately distinct from alignment: capability is how much a model can do, alignment is what it does with it.

Explainers using this term

• Designing Distillation Data — Deciding What to Ask the TeacherDeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning