In-Context Learning

Learning a new task from a few examples given in the prompt, without updating the model's weights.

In-context learning means a model performs a new task at inference time using only the task description or a few examples given in the prompt, with no gradient updates (no parameter changes) at all. OpenAI's 2020 GPT-3 paper systematically demonstrated this ability: tasks like translation and question answering could be specified just with a text description plus a few examples, and the ability grew stronger as the model scaled up. Its significance is that switching tasks no longer requires retraining, only a different prompt. In embodied AI, researchers feed a policy model a few robot demonstration trajectories as a “prompt,” letting it imitate a new task on the spot; the idea is closely related to few-shot learning, meta-learning, and prompt engineering.

ExampleICRT (In-Context Robot Transformer, 2024) feeds a few human-teleoperated demonstration trajectories — images, states, and actions — into the model as a prompt at inference time, letting a Franka arm carry out a new task without updating any parameters.

Also called

ICL

Related

Few-shot · Meta-Learning · Prompt / Prompt Engineering · Large Language Model · Next-Token Prediction · Meta Reinforcement Learning

Sources

Language Models are Few-Shot Learners (GPT-3, arXiv:2005.14165)
In-Context Imitation Learning via Next-Token Prediction (ICRT, arXiv:2408.15980)