Capability Skills
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Capability skills are a category of fine-tuning or training intervention that teach models behaviors they cannot yet perform consistently — as opposed to Preference Skills, which encode stylistic or value-based behaviors the model could perform but does so inconsistently by preference.
Definition and Core Characteristics
Philipp Schmid (Google DeepMind) introduces capability skills as one of two fundamental skill types when reasoning about what fine-tuning accomplishes. The defining criterion is inability: if a model simply cannot do something reliably at the current moment, teaching it that behavior constitutes a capability skill. 53:12
Contrast with Preference Skills
Schmid draws a sharp distinction between capability skills and preference skills on the axis of durability:
- Capability skills are inherently temporary. Because they fill gaps in what a model can do right now, they are expected to become obsolete as base models improve — the underlying gap closes, and the fine-tuned behavior may no longer be necessary or even desirable.
- Preference skills are described as "more durable" — they encode how a model should behave or express itself, which does not automatically resolve as models scale. 53:12
Implications for Evals and Lifecycle Management
Because capability skills are temporary by nature, Schmid's framing implies a lifecycle consideration: capability skills should be retired as models improve, rather than accumulated indefinitely. This has practical consequences for how teams design evals — evaluations attached to capability skills may need to be revisited or deprecated when the underlying model advances, whereas evals protecting preference skills warrant longer-term investment and protection. 53:12