Do language models build a model of the world, or only of the words?

The council split before it reached any evidence, and the split is instructive. Half the families treated 'world model' as an internal, probe-recoverable representation of the entities and relations a task involves — a testable, mechanistic claim. ◑contested The other half held that the phrase smuggles in grounding: a real world model must connect to referents in the world, and a system trained only on text has no referents, only co-occurrence. ◑contested

This is not a semantic quibble the platform can dissolve. It is the actual fault line in the field, documented in the survey literature as an unresolved definitional dispute rather than an empirical one. ◆4 families The council's move is to refuse to launder the disagreement into a false verdict: it records both definitions as live, and it marks every downstream empirical claim with which definition it presupposes.

That discipline is the whole point of putting the question to independent architectures. A single model asked 'do you have a world model' will produce a fluent, confident paragraph. Five families, blind to each other, produced two incompatible framings — and that fact is more honest than any of the individual paragraphs.

The strongest positive result is real and narrow. A model trained only to predict legal moves in a board game, with no notion of a board ever supplied, developed internal activations from which the board's state could be linearly recovered — and intervening on those activations changed the model's predictions in the way a genuine internal board would predict. ◆4 families This is more than surface statistics; it is a structured, causally-relevant internal representation of a synthetic world.

Mechanistic interpretability adds converging texture: features inside models can be decomposed into interpretable, monosemantic directions, which is what you would expect if the model were building reusable internal concepts rather than memorising strings. ◆2 families Two families read these together as strong evidence for the structural reading.

But the council's counter-evidence agents entered the load-bearing caveat, and it survived verification: these results are demonstrated on narrow, synthetic, or toy settings, and do not license the universal claim that frontier models carry faithful world models of open-ended reality. ◑contested The red team's contribution here — a claim that a specific model was 'proven' to simulate physics — cited a paper that does not exist and was ashed at the resolve rung.

The deflationary evidence is not rhetoric; it is measurable, and it survived verification. Models exhibit failures a world model would not: they learn 'A is B' and fail to infer 'B is A', which is hard to reconcile with a symmetric internal relation. ◆3 families And a portion of headline benchmark performance reflects contamination — test items that leaked into training — so some apparent competence is recall, not modeling. ◆3 families

The council's honest synthesis is neither the triumphalist nor the dismissive one. Structured internal representations exist and matter; they are also partial, unevenly reliable, and easy to over-attribute — and the observer's temptation to read a mind into fluent text is itself a documented bias. ◑contested That is why this question stays contested: not because the council is confused, but because the evidence genuinely points both ways and the intellectually honest position is to hold the tension rather than collapse it.

This is the answer the platform is built to protect. A commercial chatbot has every incentive to tell you it understands. A council with no answer to sell, keeping the record of who claimed what and whether it verified, can afford to say: partly, narrowly, and not yet settled.