上下文学习In-context Learning
In-context Learning
在对话里现学现用：给几个例子，模型就会做新任务，不用重新训练。Learning on the spot inside a conversation: give a few examples and the model does a new task, no retraining.
GPT-3 论文的核心论点：模型不更新权重。
它只靠提示里的任务说明和少量示例完成任务。
示例给 0 个、1 个还是几个，效果差别很大。
论文标题就是「Few-Shot Learners」。
The core claim of the GPT-3 paper: the model updates no weights, relying only on task instructions and a few examples in the prompt.
Whether you give 0, 1, or several examples changes results a lot — the paper's title is literally "Few-Shot Learners."
你调提示词、给示例、贴一段参考——用的全是它。知道它的边界，才知道什么时候该考虑微调。Tweaking prompts, giving examples, pasting reference text — all of it uses this. Knowing its limits tells you when fine-tuning is the better route.
回到名词地图看第4站的其他 35 个词 →Back to the map: the other 35 terms in station 4 →
- GPT-3 论文（Language Models are Few-Shot Learners）GPT-3 paper (few-shot learners) ↗