How a model learns from examples it was never trained on

The behavior, stated exactly

Put three lines in front of a model. The first reads Paris, France -> France. The second reads Tokyo, Japan -> Japan. The third reads Cairo, Egypt -> and stops, left for the model to finish.

It writes Egypt. You never told it the rule. From two finished pairs it worked out "return the part after the comma" and applied it to the third. Swap in a rule it has no name for, say "return the first three letters" or a made-up label on each line, and it still continues the pattern after two or three examples. This is the behavior to explain: a model reads a short list of input-to-output pairs in its prompt and completes the pattern for a fresh input. The field calls it in-context learning, learning from the examples sitting in the context.¹ The word "learning" is doing a lot of work in that name.

Nothing was trained

When GPT-3 first showed this off, the paper was blunt about the setup: the model is run "without any gradient updates or fine-tuning," with tasks "specified purely via text interaction."² Training is the process that sets a model's weights, the roughly 175 billion numbers inside GPT-3 that get nudged, one small step at a time, to lower its error.² When you paste your examples in, none of that happens. Not one of those numbers moves. The weights are frozen.

So whatever "learning" means here, it is not the usual meaning. It has to happen inside a single forward pass: the one left-to-right sweep in which the model reads the prompt and produces the next token, using weights that are fixed before your prompt arrives and unchanged after it leaves. The examples are input, not instruction. That narrows the question to something answerable. What is the model doing, in that one pass, with the example pairs already in front of it, that makes it continue the pattern? The answer lives in how it reads across the prompt, which is the job of attention (covered in its own lesson): each position can look back at earlier positions and pull in what they hold.

The circuit that copies from earlier in the prompt

The clearest thing anyone has found inside that forward pass is a small circuit called an induction head. It was first isolated in stripped-down two-layer models, where researchers could read the wiring end to end and watch it drive the pattern-completion behavior.³ An induction head is built from two attention heads working in series, and its job is described in one line: it finds an earlier spot where the current token already appeared, and it copies whatever came next.⁴

Walk it on a real string. Suppose the prompt contains, earlier on, the text The password is zephyr9, and now the model has read up to a second The password is and must predict what follows. Two steps run:

1

A head in the first layer is a "previous-token head." It walks along and writes onto each token a note about the token just before it. So the first zephyr9 gets tagged with "the word in front of me was is."⁴

2

A head in the second layer, the induction head proper, is now sitting on the current is. It searches earlier positions for one tagged "preceded by is," lands on that first zephyr9, and pushes up the odds that the next token is zephyr9.⁴

Researchers write the pattern as [A][B] ... [A] -> [B]: having seen A followed by B once, then meeting A again, predict B.⁴ The same shape reaches the comma example. "Return the field after the comma" is the same move: the head learns from the first lines that the useful token sits in a certain spot relative to a marker, then copies the token in that spot for the new input. The examples are not teaching the model a rule in the training sense. They are laying down earlier occurrences for a copying circuit to find. Knocking out the induction heads themselves removes the behavior, which is why researchers treat the heads as causal.⁴

The circuit shows up on a schedule

The strongest reason to take induction heads seriously is when they appear. Train a model and watch its error fall, and there is a short stretch early on where the loss dips in a small, visible bump. Right at that bump, induction heads form, and the model's grip on in-context examples jumps at the same moment.⁴ The two events are locked together across every model size tested, as long as the model has more than one layer, which fits the finding that the copying circuit needs two attention layers to exist at all.³

To see the jump you need a number for "grip on in-context examples." The measure used is deliberately crude: take the model's error on the 500th token of a passage and subtract its error on the 50th, so a model that keeps getting better the more context it has read scores high.⁴ The authors flag that the exact token positions are "somewhat arbitrary," so read it as a thermometer, not a verdict.⁴ The result has since been reproduced outside the group that found it: an independent study of a small classification setup reports that in-context learning there "is driven by the abrupt emergence of an induction head," and ties the sharpness to the order in which the circuit's parts get learned.⁵ A circuit that reliably arrives on cue, and that a second lab reproduced, is a real object.

The honest edge of the explanation

The four-word version skips one step. Everything above is solid for small models and for plain copying. Whether induction heads account for the majority of in-context learning in a large production model is not settled, and the people who found the circuit are the ones saying so. They call their own case "circumstantial" and "only the beginnings of evidence," and note the argument is stronger for small models than for large ones.⁴ An outside researcher put the limit plainly: the work is "not purporting to explain the entire mechanism," only that induction heads "seem to be involved."⁶

One finding shows why pure copying cannot be the whole account. If you take a working set of examples and scramble the labels, pairing each input with a wrong output, the model's accuracy barely drops. What the examples supply, in large part, is the format, the range of possible labels, and the kind of input to expect, not the exact input-to-label rule.⁷ A circuit that simply copied the demonstrated answers should collapse when the answers are wrong. It does not. So induction heads explain how a model continues a pattern it can find earlier in the text, which is real and common, but they do not by themselves explain how it lands the right answer on a task where the examples carry no answer to copy.

Two other accounts try to fill that space, at a higher level than individual heads. One says the pretrained model treats the prompt as evidence and infers a shared hidden topic across the examples, then predicts as if drawing from it, a kind of built-in Bayesian guess about which concept the examples came from.⁸ Another shows that a simplified attention layer can be constructed to carry out, inside the forward pass, a step of the same gradient descent that training uses, as if the model fits a tiny model to your examples on the spot, though this is shown for clean regression tasks rather than full language models.⁹ These are not rivals to the induction-head result so much as descriptions one level up: a circuit is what runs, Bayesian inference or an implicit gradient step is a claim about what that running computes. None of the three has closed the case, and a recent survey frames the mechanism of in-context learning as an open problem the field is still working on.¹