Why LLMs Can't Count Tokens: The Need for Intermediate Steps

ctjlewis · x · 2026-08-08

The author points out that Large Language Models (LLMs) struggle to accurately count characters or words without using intermediate tokens or external tools. This limitation is analogous to humans needing to count on their fingers or out loud when estimating the number of jelly beans in a jar. It is not an issue exclusive to LLMs, but rather stems from the lack of an implicit internal counting mechanism during text generation.

Original post →

More from Models

Models channel →