AI glossary · Advanced · Data & evaluation · Language models
Curriculum learning
Curriculum learning is a training method in which an AI model is given training data not in arbitrary order but in a planned sequence, from easy to hard.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Foundations & applications
- Fields
- Data & evaluation, Language models
- Added
- 2026-10-09
How it works
First, each data example is given a difficulty score. This can use human-defined criteria such as sentence length or the number of steps needed to solve a problem, or it can be based on how often the model gets that example wrong. Early in training, mostly easy data is shown, and the share of hard data increases as training goes on. The order can be fixed in advance, or the next problems can be chosen on the fly to match the model's current skill.
How difficulty changes during training
- Short, simple examples
- Mix of easy and hard
- Multi-step reasoning problems
Why it matters
If a model is given lots of hard data from the start, it can struggle to find its footing, making training slow or unstable. Building a foundation on easy data sometimes leads to faster and more stable improvement. Recently the idea has drawn renewed attention, for example by concentrating high-quality data at the end of LLM pretraining, or by picking problems in reinforcement learning that the model can only just barely solve.
Things to know
It does not always help. Many results show little difference from the ordinary approach of shuffling data randomly, so it needs to be tested experimentally for each problem and model. Badly assigned difficulty can even lower performance. Also, showing only easy data for too long can bias the model toward it, so the mix has to be tuned carefully.
Example
When training reasoning models on math and coding problems with reinforcement learning, there is little to learn from easy problems the model always gets right or hard ones it never gets right. So the training picks problems the model solves only some of the time at its current level, and switches to harder ones as it improves. This is a typical use of curriculum learning.