emergent reasoning skills

Emergent reasoning skills are abilities that are essentially absent in smaller models and then appear, sometimes quite suddenly, once a model crosses a certain scale of size, data, and compute. A small model scores near chance on multi-step arithmetic or a tricky logic puzzle; scale it up and at some point performance jumps from useless to usable, as if a capability switched on. The phrase captures the surprise that more of the same training recipe can yield qualitatively new behaviour rather than just smoother versions of the old.

How real and how sharp these jumps are is debated, and the debate is healthy. Some apparent leaps are artefacts of harsh all-or-nothing scoring; measure with a kinder, more graded metric and the curve looks smooth rather than stepped. Either way, the practical lesson stands: capabilities a model lacks today may arrive with the next scale-up, which makes future behaviour genuinely hard to predict from current behaviour. That unpredictability is part of why people study scaling so carefully and treat confident claims about hard limits with caution.