When can few-shot CoT exemplars hurt accuracy on unusual inputs?

## Exemplars are a prior, not an illustration It is tempting to read few-shot examples as an explanation the model reads and then sets aside. Mechanically it is closer to the opposite: the model is continuing a pattern, and everything consistent across the examples — reasoning length, section order, vocabulary, the set of answer values that appear, the confidence register — becomes part of the pattern being continued. That is precisely why few-shot CoT gives such good format consistency. It is also why it degrades in a specific, recognizable way when the input falls outside the region the examples cover. ## The failure, concretely A veterinary clinic triages intake with a prompt carrying four worked examples. Each demonstrates a symptom set, roughly four lines of reasoning, and a final tier drawn from the clinic's escalation ladder: routine, same-day, urgent, emergency. The examples were written from the four most common presentations. An animal arrives with a symptom combination that none of the four resembles — say, a chronic condition interacting with an acute complaint in a way the exemplars never demonstrate. Three things tend to happen at once: **Nearest-example capture.** The model latches onto whichever exemplar shares surface features and reuses its reasoning path, weighing the factors that mattered *there* rather than the ones that matter here. **Length and structure lock.** The reasoning comes out at roughly exemplar length. A case that genuinely needs more deliberation gets four lines, because four lines is what the pattern says an answer looks like. **Answer-space collapse.** The tier chosen is one that appears in the examples. If the right response was "this does not fit the ladder, escalate to a human", nothing in the prompt demonstrates that this is even an option. The output is fluent, correctly formatted, and confidently wrong — the worst combination for review, because format correctness is what a reviewer checks first. ## How you would know Aggregate accuracy hides this completely, because the tail is by definition a minority of traffic. The diagnostic move is to **slice the eval by similarity to the exemplars**: score cases that resemble the demonstrated presentations separately from those that do not. A large gap between the two slices is the signature. In production the proxies are cheap: output-length variance that is suspiciously low, an answer distribution that is a near-perfect match for the exemplar distribution, and a heavier concentration of human overrides on inputs that look unlike any example. A useful A/B is to run the same tail cases through a zero-shot variant. If zero-shot beats few-shot on the tail while losing on the head, you have confirmed lock-in rather than general task difficulty. ## What to do about it **Add an explicit escape hatch.** State in the instruction that the examples illustrate common cases only, that unusual combinations should be reasoned about from first principles, and that a specific out-of-ladder outcome exists. Demonstrating that outcome once is worth more than saying it, but even the instruction alone usually helps. Measure it — this is exactly the kind of change that feels effective and sometimes is not. **Route by input.** Send inputs that match known presentations to the few-shot prompt for its consistency, and send the residue to a zero-shot path that is free to reason and free to say "unclear". This concedes format consistency exactly where it was doing harm. **Cap the blast radius.** For a triage-shaped task, the cheapest robust design is that anything the model cannot place confidently goes to a person. Lock-in becomes a routing inefficiency instead of a clinical error. **Reconsider whether you need full traces.** Sometimes the value the exemplars were providing was format, not method. An instruction with a strict output template and no worked reasoning gives the format without the pull toward demonstrated reasoning paths. ## The judgment to show in an interview The strong answer does not say "few-shot is bad". It says the examples encode an assumption about the input distribution, names the symptoms that show the assumption has broken, and describes an eval that would surface it before a user does. Say plainly that head accuracy and tail accuracy are different numbers and that only one of them is usually being measured.