World model
An internal picture of how things work that lets a system predict what happens next.
Think of it like
Like the mental map you use to guess a rolling ball will fall off the table before it does.
Example
A model that tracks who’s holding what across a story — and correctly says the key is now in Sara’s pocket — is showing a small world model.
How it actually works
Whether LLMs truly have world models is debated: they clearly build useful internal representations of state and relationships, yet they also make errors no consistent model would. The term matters because robust planning and reasoning arguably require modeling the world, not just predicting surface text.
For product teams
Shorthand for “does it actually understand the situation, or just sound like it?” — the crux of reliability.
For engineers
An internal representation of environment state and dynamics supporting prediction; contested in LLMs.
Related
- Grounding — Tying that model to real facts.
- Representation — The internal encodings it would live in.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome