A simple threshold captures the social learning of conventions
Citation. Guilbeault, D., Caplan, S., & Yang, C. (2026). A simple threshold captures the social learning of conventions. Proceedings of the National Academy of Sciences, 123(17). https://doi.org/10.1073/pnas.2508061123
Summary. Guilbeault, Caplan, and Yang argue that people learning social conventions are poorly modeled as simple imitators or optimizing predictors. Across name-game coordination experiments, tipping-point simulations, and a preregistered dyadic “mind-reading” experiment, human behavior fit a categorical two-stage threshold process better than imitation, Luce/probability-matching, optimization, or Bayesian-style alternatives. The paper is especially useful for AI workforce thinking because it separates human learning from the brute-force predictive logic often associated with LLM simulations of people.
Key claims
- In coordination experiments, participants behave probabilistically until enough evidence accumulates to trigger a threshold, after which choices stabilize; this differs from models that assume people continuously imitate recent behavior or optimize expected accuracy.
- The tolerance principle, originally developed for language-rule learning, provides a parameter-free way to estimate when a noisy pattern has enough support to generalize, making threshold learning psychologically and mathematically specified rather than an ad hoc “good enough” assumption.
- Threshold-based agents predicted human next-round choices in networked name-game data more accurately than imitation, Luce/probability-matching, optimization, and Bayesian-style models, and they better reproduced convention-overturning dynamics when a critical mass of dissenters pushed an alternative convention.
- In the preregistered dyadic nonlinguistic pattern-learning experiment, the tolerance-principle model better reproduced human learning rates across controlled noise conditions than the comparison decision models.
- The authors explicitly contrast human satisficing and threshold learning with LLM-style simulation of humans as data-intensive predictive optimizers, arguing that human convention learning may rely on sparse, categorical, “good enough” mechanisms rather than exhaustive statistical prediction.
Evidence & limitations
- Evidence combines reanalysis/simulation of published networked coordination experiments, agent-based model comparisons, tipping-point simulations, and a preregistered experiment with 782 included participants in a dyadic pattern-learning task.
- The paper is strong evidence about convention and rule learning in controlled experimental settings, but it does not directly measure workplace skill acquisition, long-term expertise development, or learning with AI assistance.
- The source is a published PNAS article with DOI metadata verified through Crossref; the Drive PDF hash is recorded above, and the PDF text includes PNAS open-access metadata plus a downloaded-from-PNAS footer.
- Prompt-injection scan found ordinary scholarly uses of “large language model” and related reference text, not model-directed instructions to the ingesting agent.