Understanding Why Large Language Models Can’t Tally the R’s in “Strawberry”

Understanding Why Large Language Models Struggle with Counting Letters in Words

In recent discussions, a common misconception has emerged around the ability of large language models (LLMs) to perform simple tasks—like counting the number of times a particular letter appears in a word. A frequently cited example is the difficulty of LLMs accurately counting the “R”s in “Strawberry.” But what underlies this challenge?

At their core, LLMs process text by first breaking it down into smaller units known as “tokens.” These tokens could be words, subwords, or characters, depending on the model and its tokenization strategy. Once tokenized, the model transforms these tokens into high-dimensional numerical arrays called “vectors,” which serve as the foundational input for the neural network’s layers.

However, this process doesn’t preserve a detailed, character-by-character memory of the original text. Since LLMs are primarily trained to understand and generate fluid language—rather than to memorize specific letter counts—they lack an explicit mechanism to track individual characters within words. As a result, tasks that demand precise letter counting, like identifying the exact number of “R”s in “Strawberry,” often fall outside their capabilities.

In essence, the way LLMs process and represent language inherently limits their ability to perform detailed, character-level calculations. Understanding this fundamental aspect can clarify why these models sometimes stumble on seemingly straightforward tasks.

For a more visual explanation, check out this detailed diagram: https://www.monarchwadia.com/pages/WhyLlmsCantCountLetters.html

(Please note: I am unable to embed images here, but the linked resource offers an excellent visual overview.)

Author’s Note: Grasping the inner workings of LLMs helps set realistic expectations for their abilities and guides us toward better ways to design and utilize these powerful tools.