Why LLM’s can’t count the R’s in the word “Strawberry”

Understanding the Limitations of Large Language Models: Why They Struggle with Simple Counting Tasks

In recent discussions, it’s become common to see jokes about large language models (LLMs) seemingly failing at straightforward tasks—like accurately counting the number of times a specific letter appears in a word, for example, “Strawberry.” But what’s the root cause of this limitation?

At their core, LLMs process text by dividing it into smaller units called “tokens.” These tokens are then transformed into mathematical representations known as “vectors,” which serve as the model’s input for generating responses. Crucially, this process doesn’t preserve detailed information about individual characters within the input. Instead, it captures general patterns and contextual relationships at a higher level.

Because LLMs are not designed to memorize or count precise characters, their internal representations lack the granularity needed to identify how many times a specific letter occurs in a word. This means that tasks requiring exact letter counting—like determining the number of R’s in “Strawberry”—are beyond their capability, leading to seemingly simple errors.

For an illustrative diagram explaining this concept, visit: https://www.monarchwadia.com/pages/WhyLlmsCantCountLetters.html. (Please note that image posting isn’t permitted here, but the explanation on the webpage provides valuable insight.)