Showing an AI the same examples in a different order can flip nearly half its answers — sequence matters as much as content.

This explores why the order of examples, independent of what those examples actually say, can make or break in-context learning. The sharpest evidence in the corpus is almost startling: simply moving an identical block of demonstrations from the start of a prompt to the end can swing accuracy by up to 20% and flip nearly half of all predictions, with the effect holding across different task types How much does demo position alone affect in-context learning accuracy?. So brittleness isn't only about picking bad examples — position itself is a lever the model is silently pulling.

Why would position matter at all? One clue comes from how models actually generate reasoning token by token. When a model predicts each next token, it leans heavily on the tokens immediately preceding it — this 'local memorization' accounts for up to two-thirds of reasoning errors, and it worsens as inputs get longer or drift from the training distribution Where do memorization errors arise in chain-of-thought reasoning?. If the model is disproportionately anchored to whatever sits closest in the sequence, then reordering examples changes what's 'nearby' at each decision point, and the output shifts accordingly. Order isn't cosmetic; it changes the local context the model is copying from.

There's a deeper structural reason too. For sequential tasks, in-context learning only works when the context contains whole trajectories from the same setting — 'burstiness' — rather than scattered isolated examples Why do trajectories matter more than individual examples for in-context learning?. This reframes the fragility: the model isn't averaging over a bag of demonstrations, it's reading the prompt as a structured sequence and inferring the task from its shape. Break the coherent ordering and you break the signal the model was using to figure out what you want. The same lesson appears in conversation, where models lock onto premature assumptions from early turns and can't recover — order determines which guess gets cemented first Why do language models fail in gradually revealed conversations?.

If order is a liability, it's also a tool — and this is the turn most readers won't expect. Researchers have started ordering demonstrations *deliberately* to exploit the effect rather than suffer it. One approach ranks examples by how sparse their internal activations are, feeding harder (sparse) examples before easier (dense) ones like a curriculum, with no difficulty labels required Can representation sparsity order few-shot demonstrations effectively?. Another treats example selection as a formal experimental-design problem, choosing the sequence that most reduces uncertainty about the test set Can optimal experimental design improve few-shot example selection?. The brittleness and the opportunity are the same phenomenon viewed from opposite sides.

Underneath all of it sits a tension the corpus keeps circling: the model is constantly negotiating between what it learned in training and what you're showing it now. When prior associations are strong enough, in-context information gets overridden entirely, and prompting alone can't fix it Why do language models ignore information in their context?. Order is one of the few knobs that shifts that balance without retraining — which is exactly why a fragile, position-sensitive model is also a surprisingly steerable one.