Emergent Capabilities — AI Glossary

Reference details

Related terms

• Scaling Law
• In-Context Learning (ICL)
• Chinchilla Scaling

How it works

1. A capability is measured across model scales, and the claim is that it is near-absent below a threshold and present above it.
2. The apparent discontinuity depends on the metric: exact-match scoring on a multi-step task is near-zero until every step is right, which manufactures a jump.
3. Measured with a continuous metric — per-step accuracy, or log-likelihood of the correct answer — the same capability often improves smoothly across the same scales.
4. So the sharp curve can be a property of the measurement rather than of the model, which is the substance of the debate rather than a quibble.
5. The practical consequence survives either way: a capability absent at one scale may appear at the next, so a negative result on a smaller model does not predict the larger one.

Example: How a metric manufactures a threshold

A five-step task scored exact-match. At 60% per-step accuracy the whole task succeeds 0.6⁵ ≈ 8% of the time; at 80% it is 0.8⁵ ≈ 33%; at 95%, 0.95⁵ ≈ 77%. Per-step accuracy rose smoothly from 60 to 95, and the reported score jumped from 8% to 77% — a curve that looks like a threshold crossing. Nothing discontinuous happened in the model. That is why the measurement choice is part of the claim, and why continuous metrics are the test for whether an emergence is real.

Commonly confused with

Used as though it meant unpredictable, which overstates a contested finding. The documented observation is that certain benchmark scores rise sharply past a scale; the dispute is whether that reflects the model or the metric, and under continuous metrics many such curves are smooth. It is also not the same as scaling laws, which describe loss falling predictably with compute — loss can be perfectly smooth while a thresholded downstream score appears to jump, and both statements can be true of the same training run.

When to use it

Reach for it when:

• Explaining why a capability absent in a smaller model may appear in a larger one.
• Arguing for continuous metrics in evaluation, since thresholded ones manufacture jumps.
• Cautioning that a negative result at one scale does not generalise upward.

Reach for something else when:

• As evidence that capability gains are inherently unpredictable, which the metric critique undercuts.
• Interchangeably with scaling laws, which describe smooth loss rather than discontinuous capability.
• As a claim about a specific future capability, which no scaling curve supports.

Primary source

• Wei et al., Emergent Abilities of Large Language Models (arXiv)

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Emergent Capabilities — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path