Zero-Shot
Giving an instruction with no examples attached and having it work
- Zero-shot means getting a task done with no examples attached at all — the instruction alone does the job.
- It works because of pretraining. The model has already been through countless similar situations.
- The instruction is the only clue, so the clearer the verb and the conditions, the better the result.
- It holds up well on common tasks, but wobbles on a fussy format or a rule specific to one company.
- It saves room and money, so when it works, there's usually no reason to attach examples at all.
Contents
1The analogy
Ever bought a new alarm clock and actually opened the manual? Most people just start pressing buttons. Press the one marked "alarm," nudge the time with the up and down arrows, press again to save. Nobody demonstrated any of it first.
That works because this isn't a first time. Microwaves, remotes, an old clock — all of that has been handled before. Those experiences add up into a feel for "this kind of device is usually built like this." This particular clock is new; this kind of task is not.
Not everything goes smoothly, though. Setting a different alarm for each day of the week, or changing the snooze interval — pressing around rarely gets those right. They're rare enough that there's no instinct for them yet. Zero-shot looks exactly like this.
2In detail
Why no examples are needed
A model has already read an enormous amount of text during pretraining, building up its basics along the way. That text was already packed with summaries, translations, opinions written up, and labeled examples of every kind.
So a request like "cut this down to three lines" isn't being asked for the first time. What a summary looks like, what gets kept and what gets dropped, is already baked in. It's the same as reaching for experience with an old device in front of a new one.
This is the big break from the older approach. Getting a new task done used to mean gathering matching data and training on it separately. Now, a single line of instruction is enough to try.
The instruction is the only clue
With no examples, the instruction is everything. A vague one leaves nothing to lean on. That's why nailing the verb and attaching conditions to the result matters so much.
"Extract just the key points in three lines" beats "summarize this." "Pick either praise or complaint" beats "sort this." Spelling out the exact list of possible answers right in the instruction cuts down on stray results a great deal.
The expected shape can be pinned down in words too. Something like "put each item on its own line and number it" gets a format fairly well set even without an example. Worth trying before reaching for examples.
Where it holds and where it breaks
It works well on tasks the world does constantly — summarizing, translating, changing tone, sorting good from bad. Training material was overflowing with cases like these.
It wobbles on tasks with a fussy format. Item order or a delimiter has to land exactly, and no matter how carefully it's written out in words, small things slip. If a program has to consume the result, that slip becomes an error immediately.
Some things just don't work at all. A classification scheme unique to one company, a rule that changed recently, something written only in an internal document — none of that was ever encountered. That standard has to be spelled out in the instruction or handed over as material.
Deciding whether to add examples
It's worth trying with no examples first — if it works, that settles it. Examples take up room, and the more that gets sent back and forth, the more time and cost climb. Run the same task thousands of times a day and that gap adds up fast.
The signal to switch is clear: the format keeps slipping no matter how the instruction gets fixed, the tone of the answer shifts every time on the same request, or the standard is too fuzzy to put into words. That's when a couple of short examples beat a long explanation. Refining the instruction and attaching examples aren't an either-or — both can run together.
3More precisely
Zero-shot originally described solving a kind of problem never seen during training at all. For language models, it usually means giving an instruction with no examples in the prompt. There being no examples doesn't mean there was no training — pretraining and the tuning that teaches a model to follow instructions have both already happened.
The analogy breaks down somewhere. A person handling a new device can press a button, notice it's wrong, and undo it; a model doesn't try things and pick — it produces one answer in a single pass. That's why it can sound confident even on something it doesn't really know. And a person who can't make sense of a button label goes looking for a manual, while a model never asks back when an instruction is vague — it fills the gap with the most common guess instead. Getting it to ask back has to be written into the instruction too. A line asking it to say what's missing instead of answering when information is short does exactly that, and without that line, silence gets read as permission to guess.
4Try it yourself
5Common misconceptions
It's easy to think zero-shot means a model doing a task it was never trained on, but actually it works because pretraining already covered plenty of similar cases.
It's easy to think adding examples always helps, but actually if a task already works well, skipping examples is faster and cheaper.
It's easy to think a task that fails zero-shot is one the model simply can't do, but actually refining the instruction or attaching a couple of examples often fixes it.
7One-line summary
In shortZero-shot means getting a task done with the instruction alone, no examples attached, and it works best on common tasks pretraining has already covered.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02