Can adding more words to a passage actually interfere with meaning?

This explores whether adding words to a passage can actually interfere with meaning, rather than just adding harmless filler. The corpus says yes, and the most direct evidence is mechanical: reasoning accuracy collapses as input gets longer, even when the extra words are irrelevant padding and the passage stays far below the model's context limit. One study found accuracy dropping from 92% to 68% with just 3,000 tokens of stuffing, and the effect held even with chain-of-thought prompting Does reasoning ability actually degrade with longer inputs?. So length itself — not difficulty — is a degrading force.

The deeper answer is that meaning was never additive to begin with. One line of work argues that comprehension is the live detection of which subsets of words activate a shared frame — a selective, non-monotonic operation, not a sum of individual word meanings How do readers actually build meaning from words?. The mind holds frame-coherent words in tight resonance while actively suppressing linguistically adjacent but frame-unrelated ones Does the mind selectively activate frames from only some words?. If meaning is selective resonance, then every extra word is a candidate distractor: it raises the chance of activating a competing frame the reader then has to suppress. That's the mechanism behind "non-monotonic" — adding signal can subtract clarity.

There's also a structural cost. Coherent comprehension requires juggling three layers at once — the segments, the speaker's purpose, and what's currently salient — and a failure in any one disrupts the whole How do readers track segments, purposes, and salience together?. More words means more to keep aligned across all three, and more places for salience to drift. Relatedly, what predicts a reader's mental state isn't individual words but the causal reasoning that runs across statements Why do discourse patterns predict anxiety better than single words? — meaning lives in the relationships between words, which extra words can blur.

The interesting wrinkle is that not all additions hurt. Appending an emotional phrase like "this is very important to my career" to a prompt reliably improves model performance — not by adding information, but by adding motivational framing Can emotional phrases in prompts improve language model performance?. So the question isn't really word count; it's whether the added words reinforce the active frame or compete with it. Padding interferes; framing helps. And the surface form matters more than we'd like: semantically identical phrasings produce different results depending on how frequent the wording is, because the reader (here, a model) tracks statistical mass, not pure meaning Do language models really understand meaning or just surface frequency?.

What you didn't know you wanted to know: the reason "more words interfere" is the same reason concise writing feels clear — meaning is a selection problem, not an accumulation problem, and every word you add is one the reader has to decide whether to suppress.