ModelRefs / Large Language Model — AI Glossary
Large Language Model — AI Glossary
A neural network trained on massive text corpora to generate, summarize, classify, and reason about natural language. Also called LLM or Large Language Model.
Overview
Large Language Models (LLMs) are transformer-based neural networks with billions or trillions of parameters, trained on web-scale text to predict the next token. Modern LLMs power assistants such as GPT-5, Claude, and Gemini and serve as the foundation for agents, RAG systems, and copilots.
Reference details
| Topic | models |
|---|---|
| Also known as | LLM, Large Language Model |
| Last reviewed | 2026-06-24 |
Related terms
How it works
- Text is first broken into subword tokens by a learned vocabulary, which is why a model's cost and context limits are counted in tokens rather than words or characters.
- Pretraining presents a large corpus and asks for the next token at every position, scoring the prediction against what actually followed — so the supervision comes from the text itself and no labelling is required.
- That objective is what makes web-scale training possible, and it is also what the model is: an estimator of which continuation is likely, with no separate representation of what is true.
- Capability scales predictably with parameters, data and compute together, which is why model sizes and corpus sizes are chosen jointly rather than one being maximised.
- The result of pretraining is a base model that continues text but does not follow instructions, and it is a different artifact from anything deployed.
- Post-training converts it: supervised fine-tuning on demonstrations installs instruction-following, and a preference stage shapes which of several acceptable answers it gives — so an assistant is a pretrained predictor with two further stages applied.
Example: Next-token prediction, made concrete
Given “The capital of France is”, the model produces a probability distribution over the next token — “ Paris” very likely, “ a” far less so — samples one, appends it, and repeats. Everything else, including apparent reasoning, is built on that loop. It explains the characteristic failure too: a confident wrong answer is simply a high-probability continuation that happens to be false.
Commonly confused with
An LLM is not a database or a search engine. It has no lookup table to consult and no notion of whether a statement is true — only which continuation is likely. That is why grounding, tools and citation exist: they add the checks the architecture does not provide.
When to use it
Reach for it when:
- The task is language: generating, transforming, classifying or summarising text
- The input is unstructured and the rules are hard to enumerate
- An approximate answer with human review beats no answer
Reach for something else when:
- A deterministic result is required and a rule or query would give it
- The cost of a confident wrong answer is high and nothing checks the output
- The task is arithmetic or lookup — give it a tool instead
Referenced by
This term is used by the following ModelRefs references:
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Large Language Model — AI Glossary.
Frequently asked questions
What is Large Language Model?
A neural network trained on massive text corpora to generate, summarize, classify, and reason about natural language.
Is Large Language Model the same as LLM?
Yes — LLM, Large Language Model are common aliases for Large Language Model.
What concepts are related to Large Language Model?
Closely related concepts include token, context window, inference, fine tuning.