Why does keyword priming require only three training exposures to establish?
This explores why so few training exposures (just three) are enough to lock in keyword priming — and the corpus suggests the answer is that priming doesn't build new knowledge, it surfaces associations the model already carries.
This explores why keyword priming establishes after only three training exposures, and the most direct evidence reframes the whole question: priming isn't really *learning* in the sense of acquiring something new. The work on predicting priming from pre-learning probability shows that whether a keyword will become primed is already written into the model *before* any gradient updates happen — post-learning priming is strongly predictable from the keyword's pre-learning probability, with a sharp threshold around 10^-3 separating contexts where priming takes hold from those where it never does Can we predict keyword priming before learning happens?. If the association is already latent above that threshold, three exposures is enough because the training is nudging an existing pathway, not carving a new one. Below the threshold, no amount of the same exposure establishes it.
That 'activation, not injection' pattern shows up across the corpus under different names. Prompt optimization research draws the same hard line from the inference side: prompting can reorganize and retrieve what a model already holds, but it cannot supply knowledge absent from training — there's a ceiling no clever prompt can break Can prompt optimization teach models knowledge they lack?. Keyword priming is the training-time cousin of that ceiling: a few exposures activate a latent association cheaply, but they can't manufacture one from nothing.
The most striking parallel is in reasoning, where a *single* training example can lift math accuracy from 36% to 73.6% and keep improving generalization for 1,400 steps after training accuracy is already maxed out Can a single training example unlock mathematical reasoning?. The lesson is the same — when the capability is already latent, the training signal's job is to *activate* it, and activation is fast and sample-cheap. Three exposures for priming and one example for reasoning are both signatures of unlocking, not building.
Why is the latent stuff there in the first place? Other notes point back to pretraining as the layer where these dispositions get planted. Cognitive biases, for instance, are shaped almost entirely during pretraining and merely modulated by finetuning — models sharing a backbone show the same bias patterns regardless of what they're tuned on Where do cognitive biases in language models come from?. And when a strong pretrained association exists, it dominates: models will override their own context and ignore in-context information when parametric priors are strong enough Why do language models ignore information in their context?. Priming works on three exposures for the same reason context-override happens — the prior is already a deep groove, and a little reinforcement is all it takes to make it dominant.
The thing worth walking away with: the 'three exposures' number isn't a fact about how fast models learn — it's a fact about how much was already there. The threshold is a readout of latent structure, which is why you can *predict* priming before training even starts.
Sources 5 notes
Pre-learning keyword probability strongly predicts post-learning priming across architectures and model sizes, with a ~10^-3 threshold separating contexts where priming occurs from those where it doesn't. Just 3 training exposures suffice to establish the effect.
Prompting works entirely within a model's pre-existing training distribution and cannot supply domain knowledge absent from training data. This creates a hard ceiling: no prompt strategy can compensate for missing foundational knowledge, only reorganize what already exists.
A single example in RLVR boosts math performance from 36% to 73.6% and enables test accuracy to improve for 1,400 steps after training accuracy reaches 100%, revealing that minimal activation signals unlock latent reasoning capability.
A causal experiment using random-seed variation and cross-tuning showed that models sharing a pretrained backbone exhibit similar bias patterns regardless of finetuning data. Biases are planted during pretraining and merely swayed by instruction tuning.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.