“Large language models can solve the Caesar cipher for common transposition lengths like 3 or 5, but fail on arbitrary numbers like 9, indicating they have memorized specific cases rather than the general algorithm.”