LLMs are trained to predict likely language patterns, not to execute arithmetic algorithms. Simple sums like 23 + 18 appear so often in training data that models pattern-match them correctly, but exact multi-step or large-number arithmetic like 9372 × 4817 requires precise digit manipulation the model is only approximating, not computing.

Because generation is statistical token prediction rather than symbolic computation, arithmetic that's common enough in training data gets pattern-matched correctly, while arithmetic that requires exact digit-by-digit manipulation on numbers the model hasn't effectively memorized tends to drift into plausible-looking but wrong answers. The production fix isn't better prompting alone — it's giving the model a calculator or code interpreter tool and having it delegate exact computation rather than trying to generate the answer as text.