An AI stutters when counting not because it doesn't know the number, but because its brain can't find the right word for it.
Large language models represent the correct count internally with near-perfect accuracy, yet they still fail to say the right number. The internal count direction in the model's mind is mathematically separate from the part that picks which digit to say. This means the AI effectively knows it has counted five items, but it cannot read out that fact to the output head. It is a structural communication breakdown rather than a failure of logic or math. Fixing this stutter requires building a better bridge between the model's internal knowledge and its speech center.
The Right Answer, the Wrong Direction: Why Transformers Fail at Counting and How to Fix It
arXiv · 2605.03258
Large language models often fail at simple counting tasks, even when the items to count are explicitly present in the prompt. We investigate whether this failure occurs because transformers do not represent counts internally, or because they cannot convert those representations into the correct output tokens.Across three model families, Pythia, Qwen3, and Mistral, ranging from 0.4B to 14B parameters, we find strong evidence for the second explanation. Linear probes recover the correct count from