Do LLMs actually have world models or just facts?

The debate about whether LLMs develop "world models" is partly terminological. Two senses of "world model" are conflated:

Sense 1: Factual world representation. A coherent encoding of world facts — spatial relationships, temporal orderings, causal associations extracted from text. LLMs demonstrably have this — since Can large language models develop genuine world models without direct environmental contact?, they extract genuine world structure from text about the world rather than from direct environmental contact.

Sense 2: Mechanistic world model. A compact, generative model of how the world works — the kind of model that supports counterfactual reasoning, causal intervention, and novel prediction under distributional shift. The inductive bias probe evidence suggests LLMs do NOT have this: Do foundation models learn world models or task-specific shortcuts?. When tested on tasks that require genuine mechanistic reasoning (counterfactual manipulation, novel causal chains), performance collapses.

The resolution pattern: Claims that LLMs "develop world models" (Sense 1) and "rely on task-specific heuristics rather than world models" (Sense 2) are both correct. The disagreement is about which sense of "world model" matters. For many practical applications, factual representation suffices. For robust reasoning under distributional shift, mechanistic models are required.

This connects to the broader pattern of LLM capabilities that look complete from one angle and hollow from another: Can LLMs understand concepts they cannot apply?, the imposter intelligence thesis, and Can language models understand without actually executing correctly?.