Are the rare facts a model forgets during training the same ones it can't pull up when asked questions?

This question pushes on whether a single mechanism — the thinning-out of low-probability items during training — explains both what a model forgets and what a retriever can't find. The closest the corpus comes to your 'tail collapse' intuition is work showing that whether new knowledge actually sticks after a gradient update is predictable in advance from the keyword's *pre-learning* probability, with a sharp threshold around 10^-3 separating contexts where priming happens from those where it doesn't Can we predict keyword priming before learning happens?. That's a concrete tail effect: items already living in the low-probability tail before training tend to stay inert afterward, and only a few exposures are needed once you're above the line. So yes, there is a measurable 'tail' boundary in training — but it's framed as a predictor of *learning failure*, not retrieval failure.

The retrieval side of your question has its own, separate diagnosis. Retrieval failures are described as architectural rather than statistical: they happen at fixed levels — when to trigger retrieval, the mismatch between what embeddings measure (association) and what the task needs (relevance), and a hard mathematical ceiling where embedding dimension limits the set of documents that can be represented at all Where do retrieval systems fail and why?. Notice the interesting parallel: both stories end at a representational ceiling. In training, items below a probability threshold can't be primed; in retrieval, document sets beyond the embedding's dimensional capacity can't be distinguished. They rhyme — both are capacity limits, not tuning problems — but the corpus never claims one *predicts* the other.

Where the two genuinely connect is the practical fix: when the tail is the problem, sidestep the representation that collapses it. Agents that issue grep-style commands over raw text beat dense embeddings precisely on entity-constrained, lexically precise queries — the exact cases where embeddings 'conflate' rare entities into the same blurry region Can direct corpus search beat embedding-based retrieval?. That's a direct admission that the embedding space loses the tail, and lexical search recovers it. Similarly, domain adaptation can be done from a short text description alone, generating synthetic training data to pull underrepresented domains back into range without ever touching the target collection Can you adapt retrieval models without accessing target data?.

There's also a cautionary cousin to your hypothesis. Systems that replace retrieval with a single model that repeatedly compresses memory follow an inverted-U: continuous reprocessing helps up to a point, then degrades *below* a no-memory baseline through misgrouping, context loss, and overfitting Can a single model replace retrieval for long-term conversation memory?. That's tail collapse you can watch happen in real time — rare events get summarized away — and it does produce retrieval-like failure. So in the narrow case where 'retrieval' is folded into the model itself, collapse and failure really are the same event.

The thing worth taking away: the corpus reframes your question from 'does X predict Y' to 'X and Y are the same kind of limit wearing two costumes.' Both training and retrieval fail at a representational boundary — and the better predictor of trouble may not be the collapse itself but the model's own uncertainty. Calibrated token-probability uncertainty turns out to beat elaborate external heuristics at deciding when retrieval is needed Can simple uncertainty estimates beat complex adaptive retrieval?, which hints that a model often already 'knows' when it's standing on the thin part of the tail.