If an LLM recognizes letters in a word, why can it still miss repeated ones?

Answer: Counting is a separate skill

It only checks presence — Checking whether a letter is present is closer than a random guess, but it still falls short. The hard part is not noticing one r; it is reliably enumerating repeated occurrences. The cited counting work found exactly this split: recognizing letters is easier than counting them.

Tokens hide word interiors — Tokens can make word interiors less transparent, but that is not the whole answer. The sourced finding is narrower: recognizing letters and reliably counting repeated occurrences are different operations. The miss happens at the counting step, not because the written word literally lacks interiors.

Counting is a separate skill ✓ — The hard part is the counting operation, not merely knowing that a letter exists. Studies find that repeated letters and counting complexity drive errors even when models can recognize letters. The surprising split is between recognizing text and performing a tiny symbolic count over it.