What's this about?
Proposes "skill entropy," a metric for how well a model composes and deploys reusable skills across long-horizon reasoning, plus a way to benchmark and train for it (8↑).
What exactly is "skill entropy"?
Loosely, a measure of how flexibly a model draws on a diverse repertoire of skills rather than defaulting to a few. Low entropy = it keeps reaching for the same hammer; healthy entropy = it picks the right tool for each sub-step of a long task. The paper turns "uses skills well" into something measurable.
Why does it matter?
Skills are the week's throughline — the Cola store curating them, SkillJack attacking them, Emollick noting models treat them as suggestions. A metric for *how well* skills are used is the missing measurement layer under all of that.