Ch 19. Learning from examples in the prompt

Write “hot -> cold” and “big -> small” on two lines, then “happy ->”, and a model will usually answer sad. Nobody told it the task was opposites; it worked that out from the examples, without any more training. This is in-context learning, and giving a few examples like this is called few-shot prompting.^[1]

Write “hot -> cold” and “big -> small” on two lines, then “happy ->”, and a model will usually answer sad. Nobody told it the task was opposites. It worked it out from the examples, during a single forward pass, without changing a single weight. This is in-context learning, and giving a few examples like this is called few-shot prompting.^[1][1]

With no examples, Qwen2.5-1.5B-Instruct gives sad about a 1% chance; with four examples, 96%. The made-up rule works too, though less reliably.

With no examples, Qwen2.5-1.5B-Instruct gives sad about 1% for “happy ->”; with four, 96%. For the made-up rule, the right label goes from almost nothing to between 22% and 78%. Not every word gets easier: “light -> dark” became less likely as examples were added.

The much smaller GPT-2 mostly fails: it repeats the word it was given, or copies the last answer it saw. Bigger models are much better at this.^[1]

GPT-2 small mostly fails. For opposites it usually just repeats the word it was given, and for the made-up rule it answers “zork” for everything, the label of the last example it saw. In-context learning gets much better as models get bigger.^[1][1]

Answers longer than one token

GPT-2 splits the made-up labels in two: “blip” is ” bl” + “ip”, and “zork” is ” z” + “ork”. The chance of a whole answer is the product of its tokens’ chances, each given everything before it:

For “horse ->” after four examples, that’s 0.69 × over 0.99, about 0.69: once GPT-2 has written ” z”, the “ork” is close to certain. So the whole choice is made at the first token, ” z” or ” bl”, and after four examples GPT-2 picks ” z” for the fruits too.

You might expect wrong examples to teach a wrong rule. A study of 12 models found that mixing up the examples’ labels at random barely mattered: what the examples taught was which labels to use and what the answer should look like.^[2] But bigger models do follow labels that are wrong in a consistent way.^[3] Here, eight short reviews are labeled positive or negative before 24 new ones. Pick how the eight are labeled:

If examples teach a rule, wrong examples should teach a wrong one. A study of 12 models, GPT-3 among them, found otherwise: swapping the examples’ labels at random barely hurt. What mattered was that the examples showed which labels to use, what the inputs look like, and the layout of the answers.^[2][2] A later study found that size changes this. Given examples with every label flipped, small models stuck to what they had learned in training, while the largest flipped their answers too, and mainly large models could learn labels with no meaning, like foo and bar for positive and negative.^[3][3] Here, eight short reviews are labeled before 24 new ones. Pick how the eight are labeled:

Without examples, Qwen2.5-1.5B-Instruct gives positive or negative only about a quarter of its chances; with any eight examples, nearly all of them. That is the format at work. With half the labels wrong, it still labels all 24 reviews correctly. But flip every label and it follows the flipped labels for 23 of 24 reviews. GPT-2 small gets 21 right with the right labels, and with flipped or made-up labels it gives every review the same one.

Without examples, Qwen2.5-1.5B-Instruct puts only 27% of its probability on ” positive” or ” negative”, and GPT-2 almost none; with any eight examples, Qwen puts over 99% there and GPT-2 over 90%. That is the label set and the format at work. With half the labels wrong, Qwen still labels all 24 reviews by their real sentiment: labels that contradict each other teach it nothing, so it falls back on what it knows. Flip every label, though, and it follows the flipped labels for 23 of 24 reviews, and with foo and bar it uses the made-up labels correctly 23 times. In that study, it took far larger models to follow flipped labels; Qwen is a much newer model, on a much easier task. GPT-2 small gets 21 right with the right labels and 17 with half wrong, and with flipped or made-up labels it gives every review the same label.

How does the model use the examples? Part of the answer is a kind of attention head that finds earlier places where the current token appeared and looks at what came right after it. On “a b … a”, it looks back at “b”. These induction heads may do much of the work of learning from examples.^[4] GPT-2 small has one doing exactly this on the prompts above:

What does the model do with the examples? Part of the answer is a kind of attention head that finds earlier places where the current token appeared and looks at whatever came right after it. On “a b … a”, it looks back at “b”. These induction heads appear during training at the same moment models suddenly get better at learning from their context, and they may be how most in-context learning works.^[4][4] GPT-2 small has one doing exactly this on the prompts above:

Two heads make an induction head

An induction head can’t do this alone, because a token’s query and key say nothing about which token came before it. An earlier head supplies that. A previous-token head copies each token’s predecessor into it, so the position of the answer “cold” also carries a note that it came after an arrow. The induction head’s query at the final arrow then matches the positions that came after arrows, which are the answers, and its values copy those answers forward.^[4] That takes two layers of attention, one after the other, so a one-layer model can’t form one. Head 5.0 in the widget sends 98% of the final arrow’s attention to the four answers to the opposites, rising from 5% on the first to 35% on the last.

You can also just say what you want, with no examples. For a familiar task, that is often enough. Examples help most when the answer has to look a certain way, or uses labels or a style the model can’t guess. They make every prompt longer, though.

Instructions alone, called zero-shot prompting, are often enough for familiar tasks like summarizing or translating. Examples earn their place when the answer has to follow an exact format, or uses labels or a style the model can’t guess from a description, as the reviews above showed. They also cost tokens on every call, though a fixed set at the start of the prompt can be cached (Ch 26). The next chapter is about the other half: telling the model what to do.