Does generalization frequency explain why models favor upward semantic movement?
This explores whether the 'upward' drift toward abstraction in LLM outputs (specific terms getting swapped for more general ones) happens simply because general words appear more often in training — and whether frequency, not meaning, is doing the steering.
This explores whether models climb the ladder from specific to general words because general words are statistically more common — and the corpus says yes, with a clean mechanical story. The most direct evidence is a WordNet analysis showing that hypernyms (broad concepts like 'animal') simply occur more often in text than their hyponyms (specific concepts like 'pangolin') Does word frequency correlate with semantic abstraction?. So abstraction and frequency aren't two separate forces — they're correlated by the structure of language itself. The 'upward' direction is baked into which words are common.
The second half of the explanation is that models genuinely follow frequency rather than meaning. Across math, translation, commonsense, and tool-calling, LLMs reliably prefer the higher-frequency surface phrasing even when a rarer paraphrase means exactly the same thing Do language models really understand meaning or just surface frequency?. Put the two findings together and the drift is almost arithmetic: if you always reach for the more common form, and common forms are systematically more abstract, you will systematically erase specificity — the expert's precise term gets quietly traded for the layperson's generic one.
What makes this more than a vocabulary quirk is *why* models lean on frequency at all. When semantic content is stripped out of a reasoning task, performance collapses even with the correct rules sitting right there in the prompt — models lean on token associations and parametric commonsense, not symbolic manipulation Do large language models reason symbolically or semantically?. Frequency *is* the grip they have on language. So upward semantic movement isn't a bug bolted onto an otherwise meaning-driven system; it's a side effect of the only mechanism the system really runs on.
The corpus also offers a useful counterweight: strong priors override what's in front of the model. Even when context supplies the specific, correct term, models can default back to their dominant training associations, and prompting alone won't fix it — you need to intervene in the representations themselves Why do language models ignore information in their context?. That suggests the upward pull isn't fully steerable by asking nicely for precision; the frequency gradient is wired into the weights.
The thing you might not have known you wanted to know: this means abstraction drift is a *predictable, measurable* property, not a mystery. Because hypernym-over-hyponym frequency can be counted in advance, you can in principle anticipate which specific terms a model will dissolve into vagueness before it ever generates a token — the same way pre-learning probability predicts which keywords will prime after training Can we predict keyword priming before learning happens?.
Sources 5 notes
WordNet analysis shows hypernyms (general concepts) occur more frequently than hyponyms (specific ones). Combined with LLMs' frequency bias, this means preferring common paraphrases systematically drifts toward abstraction, erasing expert-level specificity.
LLMs show consistent preference for higher-frequency surface forms over semantically equivalent rare paraphrases across math, machine translation, commonsense reasoning, and tool calling. This suggests models track statistical mass from pretraining rather than meaning-recognition as their primary mechanism.
When semantic content is decoupled from reasoning tasks, LLM performance collapses even with correct rules in context. Models rely on parametric commonsense and token associations rather than formal logical manipulation, constraining reasoning to training distribution semantics.
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.
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.