What are emergent abilities in large language models?
They are abilities that are absent in small models and present in large ones. The standard definition comes from Wei and colleagues: “We consider an ability to be emergent if it is not present in smaller models but is present in larger models” (Emergent Abilities of Large Language Models, TMLR 2022, read 16 November 2026). The same abstract states the practical consequence: such abilities “cannot be predicted simply by extrapolating the performance of smaller models.” A capability test on a small model tells you little about what a larger one will do. That is one reason each new frontier model needs fresh evaluation rather than assumptions inherited from the last one.
Are emergent abilities real or a measurement artifact?
Partly an artifact, according to the strongest counterargument. Schaeffer, Miranda and Koyejo argue that emergent abilities “appear due to the researcher’s choice of metric rather than due to fundamental changes in model behavior with scale”: nonlinear or discontinuous metrics produce sudden jumps, while linear or continuous metrics on the same outputs show smooth, predictable improvement (Are Emergent Abilities of Large Language Models a Mirage?, 2023, read 16 November 2026).
| Emergence view | Metric artifact view | |
|---|---|---|
| Claim | New abilities appear abruptly past a scale threshold | Jumps come from all-or-nothing scoring |
| Cause | Changes in model behavior with scale | The researcher’s choice of metric |
| What to do | Expect surprises when you upgrade | Score partial credit to see change coming |
For engineering, both readings lead to the same practice: measure the model and metric you will actually ship. Score partial credit as well as pass or fail, and keep test items out of training data, the problem benchmark contamination describes, so a jump you see is real.
What does emergent behavior look like in AI agents?
In agent systems, emergence comes from interaction rather than scale: simple agents following their own rules produce group behavior none of them was given. The pattern shows up across multi-agent systems:
| System | What emerged | Source |
|---|---|---|
| 25 generative agents in a simulated town | From one agent’s wish to throw a Valentine’s Day party, others spread invitations, asked each other to it and arrived together | Park et al., 2023 |
| Hide-and-seek agents trained with reinforcement learning | Six phases of strategy, including shelters built from boxes and ramps used to get past them | Baker et al., 2019 |
| Anthropic’s multi-agent research system | Early versions spawned 50 subagents for simple queries | Anthropic, 13 June 2025 |
All three read 16 November 2026. Anthropic’s engineers put it plainly: “Multi-agent systems have emergent behaviors, which arise without specific programming.” In production, emergent behavior is mostly a reliability problem, not a marvel. The mechanism that organizes a party also produces runaway delegation, agents feeding each other’s errors, or a shortcut that games the objective, the failure reward hacking names.
How do you test for and contain emergent behavior?
Evaluate the system, not only its parts. Testing each agent alone misses behavior that appears only when they interact, so run end-to-end AI agent evals on realistic tasks and read the agent trajectory, not just the final answer. Put hard limits on what can compound: subagent counts, turns, spend and tool permissions. Treat every model upgrade and every new agent as a change that can create behavior you have not seen, and rerun the evals before shipping. Red teaming catches emergent failures that scripted tests miss, because it hunts for them on purpose.