The World Model
A large language model never sees the world directly — it receives only words, that is, descriptions left behind by human experience. The fundamental question is whether, in learning to guess the next word of a story, the machine comes to build inside itself a real “map” of the things it speaks about. On a simple game board we can test this hypothesis directly, with the figures on the table: does the machine know where the pawn is, and does it rely on that internal map when it takes decisions?
Terms: a world model hidden state a probe a control causal intervention
The shadow and the real thing
All a language model has ever received is text: long strings of words set one after another. But text is not the world itself. It is only a projection of it, a “shadow” cast on paper or on a screen by people who lived through experiences, observed facts and chose what was worth telling. Like the shadows on the wall of a cave, text loses a great part of the colours, distances and depth of physical reality.
In the lessons on language we saw that, to predict the next word of a sentence successfully, the model must master a great many things: grammatical agreement, notions of common sense, relations of cause and effect, and threads of narrative. From this a fascinating hypothesis is born: might not the most efficient way for a machine to predict the shadow be to reconstitute, at least partly, the object that casts it? A model that keeps an internal record of the objects and relations in space will make far better predictions than one that merely tries to memorise statistically which words followed which in books.
This hypothesis need not be taken on trust — it can be tested experimentally. In this lesson we shall build a small world in miniature, simple and transparent enough for us to look directly into the machine’s internal memory and see whether it has drawn itself a map of reality.
A world seen only through its traces
Our test world is a simplified chessboard of 4×4 squares. On it we place a pawn that starts from the top-left corner and takes 10 steps in a row. At every step the pawn chooses at random, with equal chances, one of the cardinal directions (North, East, South, West), with a single common-sense restriction: it may not jump over the edges of the board.
Our artificial model does not see the board at all, nor the coordinates of the squares. It receives only the string of moves, like a story made of four words (“East, South, East, North…”), with a single task: to guess the next move. The network is a recurrent one, which at every step updates an internal memory made of 24 numbers (the hidden state).
The task hides a strict logical requirement: to make correct predictions, the model must give zero probability to moves that would take the pawn off the board. But to know which moves are forbidden, the machine must know which square the pawn is standing on at that moment. The position is never revealed to it directly; it can only be deduced and reconstituted step by step from the history of the moves so far.
The bars on the right show the model’s probability distribution for the next move. The red bars mark the directions that would take the pawn off the board. Notice how, after training, these forbidden moves receive nearly 0% chance, wherever the pawn has got to and however tangled the route was. Although we never showed it the board, the model behaves as if it knew exactly where it stands in space.
Can the position be read from the internal memory?
That the model “behaves as if it knew” is not proof enough. We want to know for certain whether the pawn’s position is actually written in the 24 numbers of its internal memory. To find out, we use a a probe — a separate classifier, like a doctor’s stethoscope, trained to read the internal state of the network (without changing it) and to guess which of the 16 squares the pawn is on.
To validate the result rigorously, we compare the probe with two controls: 1. The surface control: how well we could guess the position by looking only at the last move and the step number (the information visible directly on the surface). 2. The untrained network: the same architecture with random weights, to check whether any mechanical mixture of numbers might not keep some residual trace of the steps.
The result is categorical: from the memory of the trained network the probe reads the exact square with huge precision, far outstripping both controls. Nobody ever asked the network to memorise the pawn’s position. It was asked only to predict words, and the spatial representation of the square appeared spontaneously in memory because it was the shortest and surest road to a successful prediction.
Does the model use this representation?
That a piece of information can be read from a network does not yet guarantee that the network relies on it. The pawn’s position might have sat in memory as a passive echo only, without influencing the model’s real decisions. To demonstrate a direct causal link, we need a intervention.
The procedure is clean and elegant: we take the model’s internal memory at some moment and, following the direction identified by the probe, we artificially change its numbers — mentally “teleporting” it to another square of the board, one with different edges and different permitted moves. We leave everything else unchanged and ask the model for its prediction. If the model really leans on the position in memory, the moves it considers permitted should instantly become those of the new square, not of the one where the pawn actually is!
As a control, we apply a push of the same magnitude but aimed in a completely random direction.
When we shift the state towards the target square, the prediction follows faithfully: the model considers permitted the moves specific to the new square, ignoring where the pawn actually is. The random push, by contrast, merely confuses the network without steering it anywhere. This demonstrates beyond doubt that the internal position is no mere side effect: it is the active reference the model uses in order to decide.
This is the essence of what we call a a world model: a property of the outside world, deduced solely from textual traces, written coherently into the internal memory of the system and used actively in generating answers.
What this experiment tells us and what it does not
Our world in miniature had only 16 squares, and the model had 24 numbers at its disposal to remember them. A large language model faces a boundless world, has billions of parameters at its disposal, and human texts are infinitely more nuanced than a mechanical string of cardinal moves. What we have shown here is that the mechanism works: a neural network has the capacity to synthesise the geometry and the rules of the world starting from words alone.
Similar experiments carried out on frontier models — tested on games of chess, maps of cities or the execution logic of software programs — have confirmed that probes really do find coherent internal maps, and interventions prove that the machines use them.
Still, it is essential to understand the limit of this finding. An internal model of the world does not mean a direct anchoring in reality. The pawn’s position was deduced from text and stays captive in text: if we had sent the model a string of imaginary moves from an absurd world, it would have built a map for that fiction just as obediently. The phenomenon of confabulation (or “hallucination”) remains a reality, because the machine has no physical senses with which to check the world directly.
What we do gain is a deep change of perspective on the question “what does a machine know”: not merely whether it memorises associations of words, but whether it manages to build structured representations of reality from which to draw useful inferences.
In the next lesson we explore what it means for such a machine to become “cleverer”, and why its progress is often far more continuous than it looks on paper.