Presenting training material in a deliberate order, easy things first, rather than at random. It mirrors how people are taught and the evidence for it is mixed: it helps in some settings and makes no difference in many. Where it does matter for large models is in what gets shown at the end of training.
Technical view
Orders examples by a difficulty measure, either fixed or self-paced by current model loss. Results in supervised learning have been inconsistent. For large language models the effective version is data mixture scheduling, upweighting high-quality and domain-specific data during the final phase of pre-training, which reliably improves benchmark performance and is now standard practice at the frontier.
Related
Part of AI Terminology Index, section: Training and Post-training.