A common point of contention in discussions about Large Language Models (LLMs) like ChatGPT revolves around the notion of a “world model”: specifically, whether LLMs acquire something like a robust world model through their training process, or whether they fail to do so—perhaps because of inherent limitations in their design.

World models are in the news these days, though people don’t always agree in how exactly to define them, or how to determine whether a system like an LLM “has”1 one. But most definitions—like Melanie Mitchell’s in this article—do emphasize the role of causal abstractions. The world around us is vast and complex, and rather than reconstructing it atom-by-atom (see Borges’s On Exactitude in Science), it’s probably more useful to compress (or “forget”: see Borges’s Funes the Memorious2) certain details in the interest of building mental representations that can guide our future behavior.

This is, of course, a familiar concept to Cognitive Science. Cognitive scientists don’t always agree on which abstractions organisms form or how they come about, but the basic idea of a representational abstraction or “internal model” has long played a central role in explanations of the mind and brain. Although some researchers do reject the notion of mental representations altogether,3 I’d argue the dominant paradigm is one that views the mind/brain as constructing internal models of some kind.

All of which is to say: debates about “world models” are very familiar to me, and are occasionally frustrating because of a lack of specificity in what, exactly, is meant by the term. In cases like this, I think theoretical specificity is a virtue. Any specific philosophical commitment is almost certainly wrong—in the sense that it must, by definition, neglect certain details—but making (or rejecting) commitments is important for getting clarity, which in turn is crucial for determining what we think and also how to go about testing it empirically.4 As Max Weber argued, clarity is perhaps the best science has to offer.5

My entry point into this question actually has little to do with Artificial Intelligence. When I started graduate school, I was interested in narrative construction and comprehension: what happens in our minds while we read or listen to a story? I’d been reading narratology (e.g., Vladimir Propp’s Morphology of the Folktale), and while there was something appealing about the idea of a “grammar of narrative”, I wanted something more cognitively inspired, which is what led me to the research literature on situation models. Ultimately, I ended up focusing on different (though related) topics in my PhD, but that literature has stayed with me through the years, and I think it contains lessons for contemporary debates about AI world models.

Memory constraints and the power of inference

Humans don’t have infinite memory. When we read or listen to a story, it’s quite hard to remember a specific word we encountered five sentences ago. Thus, linguistic input faces what some researchers call a “now-or-never bottleneck”. To deal with this bottleneck, the mind must rapidly compress that input into some kind of (possibly hierarchical) representation that allows us to capture the “gist” of what was said while also forgetting some of the finer-grained details that may not be crucial for comprehension.

According to this account, memory constraints induce the need for representational abstraction. But which abstractions? At some level, a useful abstraction probably includes a representation of the event or “scene” described in language. This is sometimes glossed as the following question: who did what to whom? It’s no accident that language contains structural regularities that allows comprehenders to quickly extract this information from a sentence: “the lion ate the man” is, after all, very different from “the man ate the lion”.

Moreover, humans often “go beyond” what’s said. These inferences are fundamental to language comprehension. Consider, for instance, the following text:

I went to my friend’s birthday party last night. I had a hard time waking up this morning!

A comprehender likely infers a causal connection between these events: the speaker had a hard time waking up because they went out the night before. Moreover, you might further infer the specific cause: perhaps the speaker had too much to drink at the party. These inferences could in principle be wrong—they’re not logically entailed by the text—but they’re also pretty reasonable. Moreover, some inferences, like the causal connection between the events, feel so obvious that we’d likely balk if the speaker subsequently insisted the events were unrelated.

Once you’re looking for them, you notice these inferences everywhere. So much of the meaning we extract from language is not directly in the text: it’s something we construct by virtue of connecting the specifics of what we read or hear to our background knowledge and expectations. These expectations in turn shape the inferences we make. Returning to the example above: if you know the speaker doesn’t drink, then you probably won’t infer that they over-imbibed the night before—you might instead assume that they simply stayed up too late.

Linguists and psychologists have identified all sorts of inferences that underlie successful communication. Pragmatic inferences have to do with inferring a speaker’s intended interpretation based on what they did and didn’t say, as with scalar implicature: “Some students passed the test” suggests not all students passed, even if the all interpretation isn’t semantically impossible. Instrumental inferences have to do with inferring the specific instruments involved in an event: if we hear “John shaved this morning”, we probably assume John used a razor—if John had instead used something very unusual (like a steak knife), the speaker likely would have specified this. In each case, we bring to bear certain assumptions about how communication works and the ways in which meanings are typically expressed.

None of this tells us what kinds of abstractions humans form. But it does give us some clues. A situation model should be able to accommodate at least two facts: first, abstractions must be formed relatively quickly, in ways that allow us to capture the overall “scene”; and second, the abstractions should allow comprehenders to draw inferences about what wasn’t said—and these inferences, in turn, should influence the content of the situation model.

So what goes into a “situation model”?

One account comes from cognitive scientist Rolf Zwaan, who proposed the event-indexing model. According to this framework, the role of a situation model is to monitor and represent information about events and the actions of entities involved in those events. In a 1995 article, Zwaan and his co-authors suggested that this model might include at least five types of indices: temporality (an event’s time frame), spatiality (an event’s spatial region or location), protagonist(s) (an event’s key “players” involved), causality (an event’s causal relationship to other events), and intentionality (an event’s relationship with the goals of various protagonists).

Consider, for example, the dimensions involved in the following sentence:

Sam signed a lease for an apartment near the Blue Line so he could easily take public transportation to UC San Diego.

We have at least one (named) protagonist: Sam. We also have a general spatial location: an apartment near the Blue Line, one of the routes on the San Diego light rail (though notably, the signing of the lease need not have occurred in this location!). The verb’s in the past tense (“signed”), so it suggests the event already took place. And intentionality is also explicit: Sam’s motivation is being able to easily take public transit to campus.

As language unfolds, comprehenders dynamically update their situation model according to changes along each of these indices. Zwaan et al. (1995) write:

Then the reader monitors whether incoming story events require updating an index on any of these situational dimensions. For example, if a clause indicates a time shift compared with the previous clause, then the temporal index of the model needs to be updated. When an incoming event takes place in a different spatial region, the spatial index needs to be updated. When an incoming event involves a different protagonist, the protagonist index needs to be updated. When an incoming event is causally unrelated to the previous event, the causal index needs to be updated. Finally, when an incoming action introduces a new goal structure, the motivational index requires updating.

There’s a lot this initial framework leaves out: for instance, it doesn’t tell us precisely how these indices are identified or updated.

But it does offer a specific, testable theory of what goes into a situation model. The theory is also intuitively plausible, at face value: it makes sense that comprehenders would track these dimensions, given both our own phenomenology of understanding language and also what we know about language structure itself, which contains a number of explicit markers relevant to these indices.

Perhaps more crucially, the model satisfies the two requirements we enumerated above: first, these five dimensions seem like good criteria for the “gist” of linguistic input; and second, they provide the bedrock for future inferences and can themselves be informed by these inferences (e.g., inferring a causal connection between events).

Situation models: the details

From one perspective, the question of whether people form situation models feels quite obvious. It’s hard to argue with the claim that language comprehension consists in part of tracking information about events—such as who was involved, as well as when, where, and why it took place. Moreover, if someone is asked an explicit question about these features, they can probably produce a reasonable answer, provided they attended to the text in question. One might even argue that forming these situation models is constitutive of successful comprehension.

But as always in Cognitive Science, the details matter. For instance: when do people construct these situation models? Perhaps there are some circumstances where comprehenders build rich representations of a text, and others where they extract a relatively shallow or “good-enough” representation. Moreover, do people construct these models automatically while reading/listening to language, or are they built more strategically in a post-hoc manner—perhaps reflecting the context-dependent needs of the comprehender?

You might also wonder whether certain features are more salient than others. Do people attend more to the agents involved in an event than the spatial location of the event? Is there legible variance across individuals or contexts in this tendency?

Additionally, you might ask about the representational format of these situation models. Are they best described as “symbolic” in nature? Or are they more like grounded representations of an event—a kind of “mental simulation” of the scene and its protagonists, as described by language? And again, does this vary across individuals or the kind of scene being described?

These questions and others have been the focus of research on situation models for multiple decades (see Zwaan’s 2025 review paper for a summary of past and current directions). This work has used a variety of methods common to Psycholinguistics—reading time studies, eye-tracking, EEG, verb clustering, and more—to investigate the dimensions along which comprehenders reliably form abstractions while reading or listening to language. Much is still unknown, of course; but in my view, one lasting and important contribution is actually the framework itself—concretizing a “situation model” in terms of specific event features that can subsequently be operationalized and tested.

LLMs and situation models

I started with the observation that many contemporary debates about LLMs revolve around the question of whether these systems form “world models”. This question is often construed as foundational to the deeper question of whether (and what) LLMs understand or even whether and to what extent LLMs are intelligent.

In some cases, these debates are purely theoretical. Critics might suggest that LLMs are by definition incapable of constructing world models because of how they’re trained (e.g., predicting tokens in a string) or designed (e.g., neural networks without explicit symbolic representations). Here, the suggestion is that something like a world model is either implausible or impossible in such a system. In turn, proponents might respond by arguing that the training signal for many contemporary LLMs is typically much richer than characterized by critics (e.g., using reinforcement learning with verifiable rewards), and that even learning to predict text tokens encourages a system to “reverse-engineer” the generative process giving rise to those tokens, i.e., a kind of world model. In response to the point about symbolic representations, they might suggest that this is not actually a limitation: after all, cognitive scientists don’t universally agree that the human mind is best described in terms of explicit, propositional representations either—there’s a long tradition of scholars describing the mind as a continuous state-space.

Some of these debates recruit empirical evidence as well. For example, skeptics might point to common errors in planning or spatial reasoning. Anecdotally, I’ve noticed that even commercial LLMs like ChatGPT sometimes struggle with route planning, mixing up cardinal directions or suggesting a “short drive” to locations that are multiple hours away.6 That said, even “base” LLMs demonstrate a surprising ability to recall facts, answer comprehension questions about passages of text, display “commonsense” knowledge about the world, and solve Theory of Mind tasks.7

Where does that leave us?

My point in briefly laying out “both sides” here is not to suggest that both are equally plausible. Rather, my goal is to illustrate that the theoretical and empirical cases typically marshaled are often of a somewhat ad hoc nature—unanchored by a precise, testable framework for what constitutes a coherent “world model”. In the absence of such a framework, it is easy for arguments to go in circles: any piece of evidence can be seized upon or dismissed as needed because the inferential stakes have not been determined.

As I argued earlier, this is why I think specificity is a virtue. While the event-indexing model is by no means the only plausible framework of what it means to understand language, it does offer specific, testable criteria for what constitutes a situation model. Its specificity also leaves it open to correction and revision—a crucial part of any theory. In my view, research on LLMs and whether (or what) they “understand” would do well to adopt such a framework as a theoretical anchor. It will, at any rate, likely be a central topic of my coming research.