How do few-shot demonstrations signal the label space, and what happens to unseen classes?

## Label space is a format property, not a content one It is tempting to think of the label as the *answer* to each demonstration and therefore a matter of correctness. Operationally it behaves more like part of the format: the set of strings that occupy the label position defines the vocabulary the model draws from when it fills that position for the live query. Ablations of in-context learning make the same point from the other side — replacing labels with strings from outside the task's label set damages accuracy far more than shuffling correct labels within the set. The set is doing work that the individual assignments are not. ## The unseen-class failure Consider a shipment-exception classifier with eight outcomes in its schema: `damaged`, `misrouted`, `delayed_weather`, `delayed_carrier`, `address_invalid`, `refused`, `lost`, `customs_hold`. The instruction lists all eight. The exemplar block, assembled from whatever real tickets were handy, contains six — the two rarest never made it in. What you observe in production is not an error. It is silence: `customs_hold` cases come back as `delayed_carrier` or `misrouted`, plausibly and confidently, and nobody notices until a downstream team asks why the customs queue is empty. The class is available in the instruction, but the demonstrations supply no evidence that it is ever the answer, and the nearest demonstrated neighbour absorbs it. This is a coverage problem with a coverage fix. Treat the exemplar block as a matrix: schema classes down one axis, exemplars across the other, and require at least one cell filled per class. That check is cheap, mechanical, and catches the defect at authoring time rather than in a quarterly audit. It also has to be re-run whenever the schema grows — adding a ninth class to the instruction without adding an exemplar for it produces exactly the same silence. ## Exact strings and the verbalizer The label position teaches spelling as well as membership. Three practical consequences: - **Casing and separators must be consistent.** If two exemplars say `customs_hold` and one says `Customs Hold`, expect both forms in the output, and expect whatever consumes the response to be the thing that discovers it. - **The words themselves matter.** The natural-language surface of a label — the *verbalizer* — carries meaning the model can use. `refused` and `address_invalid` are semantically informative; `class_4` and `class_7` are not. Opaque codes throw away the model's prior knowledge and force the demonstrations to teach the entire mapping from scratch, which is precisely the situation where exemplar labels stop being robust to noise. If your downstream system needs numeric codes, have the model emit the meaningful string and map it to the code in your own code. - **Near-collisions are a real failure source.** `delayed_weather` and `delayed_carrier` share a prefix and a concept. If the demonstrations do not make the discriminating feature visible in the input text, the model has nothing to separate them on and the split will be close to arbitrary. ## The demonstrated distribution is a prior Beyond membership, the *proportions* in the exemplar block are read as information about how often each label occurs. A block with nine `routine` examples and one `escalate` signals that escalation is rare, and borderline cases will drift toward `routine`. This is a lever, and it points in a direction that depends on your costs: if a missed `customs_hold` is expensive and a false one is cheap, an exemplar block that mirrors true production frequency — where the class is 0.5% of traffic — is the wrong design. Over-representing the rare, expensive classes relative to the wild distribution is a legitimate and common choice. The honest caveat is that the exemplar mix is a crude control. It interacts with everything else in the prompt and is not a calibrated dial, so the mix should be validated on a held-out set with the class distribution you actually care about, not reasoned about from first principles. ## Open-ended outputs Not every task has a finite label set. When the output is free text — a summary, a rewritten paragraph, an extracted list — the demonstrations still define an implicit output space: the length, the register, the vocabulary, the level of hedging, whether units are included, whether the model says "unknown" or guesses. Everything in the label-space argument transfers. The most useful special case is teaching abstention: if none of your demonstrations shows the model declining to answer, it will not decline, and you have effectively removed "I don't know" from the output vocabulary. Where a wrong confident answer is costly, at least one exemplar should demonstrate the abstention, in the exact form you want it.