What language do large language models understand?

What language do large language models understand?

Large language models can process human languages, but their internal calculations operate on numbers. A sentence is converted into tokens, token IDs, and numerical representations before the model computes a response. Saying a model “understands English” describes its behavior; it does not mean there is an English-speaking person inside the software.

From a sentence to numbers

Imagine entering “Please explain rain.” A tokenizer divides that text into units from its vocabulary. Those units might correspond to whole words, word fragments, punctuation, or bytes. Each unit receives an integer ID. The exact division depends on the tokenizer, so this sentence has no universal token count.

An embedding layer then maps each ID to a learned vector. Subsequent neural-network layers combine information across the sequence, creating representations that reflect context. The word “rain” therefore participates in a different computation from “train,” even though their spellings overlap.

Learning language patterns

Training adjusts model parameters so the system becomes better at its learning task, often predicting subsequent tokens. Across many examples, it can acquire patterns involving grammar, factual associations, style, and relationships between ideas. These abilities vary across languages and subject areas because training coverage and model design vary.

Returning a readable answer

For a typical text-generating LLM, the network produces scores for possible next tokens. A decoding strategy selects a token, and the process continues. The tokenizer converts the resulting IDs into readable text. Numbers enable this computation, but numerical processing alone does not guarantee a correct explanation or human-like comprehension.