Zero-shot learning is the ability of a model to carry out a task correctly having seen no labeled examples of that specific task, either during training or in its prompt. In the context of large language models, zero-shot use typically means giving the model only an instruction, such as "translate this sentence into French" or "classify this review as positive or negative," and expecting a correct response with no demonstrations attached. The term predates large language models and was originally used in computer vision and classical machine learning to describe classifiers that could recognize categories never present in their training data, usually by relating a new label to known ones through auxiliary information such as text descriptions or attributes.
In language models
Modern zero-shot performance in language models rests on two developments. First, large-scale Pretraining on broad text corpora exposes a model to an enormous range of task formats, phrasings, and genres, so that a novel instruction often resembles something the model implicitly learned to do during training even without a dedicated example. Second, instruction-tuning and alignment techniques such as RLHF explicitly train models to follow natural-language instructions rather than only to continue text in a stylistically plausible way, which is what made conversational assistants such as ChatGPT usable directly from a plain-language request. Before this second step, base GPT-3-style models could often be coaxed into a task zero-shot, but reliably and helpfully following an arbitrary instruction without any examples improved considerably once instruction-tuning became standard practice across the industry.
Zero-shot versus few-shot
Zero-shot learning is usually discussed alongside Few-shot learning and In-context learning as one point on a spectrum defined by how many task demonstrations appear in the prompt: zero for zero-shot, a handful for few-shot. The GPT-3 paper reported that few-shot prompting generally outperformed zero-shot prompting on the same base model, especially for tasks with an unusual output format, but the gap has narrowed for well-aligned chat models on common tasks, since those models are trained specifically to interpret and act on unadorned instructions. Zero-shot prompting remains the default mode of interaction for most consumer chatbot use, since asking a user to supply examples before every request would be an unreasonable interface burden.
Chain-of-thought and zero-shot reasoning
A notable extension is zero-shot Chain-of-thought prompting, in which appending a simple phrase such as "let's think step by step" to a zero-shot instruction measurably improves accuracy on arithmetic and logical reasoning tasks, without providing any worked examples. This finding, reported by researchers in 2022, showed that reasoning ability elicited through few-shot demonstrations in earlier work could often be triggered zero-shot as well, simply by prompting the model to externalize intermediate steps, and it fed directly into the design philosophy behind later dedicated Reasoning model systems.
Limitations
Zero-shot performance is uneven: it tends to be strongest on tasks that are common in natural language (summarization, translation, sentiment classification) and weaker on tasks requiring precise, unfamiliar output formats or domain-specific conventions the model has little exposure to. Ambiguous instructions can also be interpreted differently than intended without examples to anchor the model's understanding, which is one reason Prompt engineering practice often recommends adding at least a few demonstrations, a system message, or explicit formatting rules for tasks where reliability matters. Zero-shot capability is also frequently used as an evaluation lens in academic benchmarks, since it more directly measures a model's generalization from pretraining than few-shot scores, which can be inflated by examples that happen to closely match the test distribution.