Zero-Shot Learning
Học không cần mẫu (Zero-Shot)
Asking an LLM to perform a task without providing any examples, relying entirely on the model's pretrained knowledge.
Zero-shot learning means asking the model to do something it has never been explicitly shown, relying on its general pretrained knowledge and instruction-following ability. The "zero" refers to zero examples provided in the prompt.
**Basic example**
"Classify the sentiment of this review as positive, negative, or neutral: 'The product stopped working after two days.'"
No examples provided — the model applies its understanding of sentiment and the labels you specified.
**Why large models are good at zero-shot**
Frontier LLMs are trained on so much text that they've implicitly learned most common tasks. They've seen thousands of sentiment analysis examples, translation tasks, summarization tasks, etc., in their training data. Instruction-tuning (RLHF) further sharpens their ability to follow new instructions without examples.
**Zero-shot vs few-shot**
Zero-shot is easier to set up but typically less accurate, especially for unusual formats or niche tasks. Few-shot is more reliable when you have clear examples of what you want but costs more tokens.
**Zero-shot chain of thought**
Combining zero-shot with chain-of-thought ("think step by step") is one of the most effective and easy-to-use techniques for complex reasoning tasks. No examples, but the model is nudged to reason before answering.
**When zero-shot fails**
Tasks with unusual output formats, domain-specific jargon, or ambiguous instructions often need few-shot examples or fine-tuning. If your zero-shot outputs are consistently wrong in the same way, that's a signal to add examples.