curriculum learning

A training strategy that controls what the learner sees when, starting with easy examples and moving to harder ones or gradually shifting the data mixture toward the target task, instead of sampling uniformly at random. Difficulty can be defined by hand (shorter sentences, cleaner images, simpler math problems), by a teacher model's loss, or by the learner itself (self-paced learning admits examples as its own loss on them falls); automatic curricula adapt the mixture online, for example by sampling tasks where learning progress is fastest. Its namers, Yoshua Bengio and colleagues, expected faster convergence and, on non-convex problems, better optima, as with continuation methods that solve a smoothed problem first. Evidence is mixed: careful studies of standard supervised benchmarks often find little gain over random ordering, but curricula matter in reinforcement learningreinforcement learning, n.: A way of teaching a computer program by trial and error: it tries actions, earns points for good results and penalties for bad ones, and gradually learns which choices pay off over time., where agents may never see a reward on hard tasks (sparse rewardssparse reward, n.: A situation in which an AI agent gets feedback only rarely — say, only when it finally wins — so it has almost nothing to learn from until it stumbles on success.), and in language-model training, where late shifts toward higher-quality or harder data (annealingannealing, n.: The last stretch of training, when the model's learning is slowed to a stop — often while it is fed a small helping of the best data — so it settles into its final, sharpest version., mid-trainingmid-training, n.: A middle stage in building a language model, between its broad initial training and its tuning into an assistant: it keeps learning the same way, but from a smaller, carefully chosen diet of math, code, reasoning or long documents.) and difficulty-filtered problem sets in RL post-training act as curricula.