Agreed! We talk about those briefly in the paper. They engage the language system strongly (b/c they place demands on ling. computations), but probably aren't exported because there's no real meaning being constructed. Where would such a representation go?

That's an interesting q. My guess would be that we get by with shallow understanding for a lot of these conventionalized metaphors. Imagine your lang. system is learning the statistical regularities for two (prob. more) senses of "breeze": one literal, one metaphorical. Cool to think about though!

Thanks Elliot, this is all very interesting. I agree that drawing connections across Marr's levels is difficult and something that we should be trying to do more of.

Thanks for the reference! Will take a look. Agreed that this interface is very important--especially now--for understanding both natural and artificial intelligence.

In summary, we offer a conceptual framework for how to think about language comprehension–beyond the parsing of sentences and the construction of shallow meanings—to advance a richer theory of what it means to understand language cognitively, neurally, and computationally.
13/13

2️⃣ Akin to deep understanding in brains, LLMs may process language more deeply by exporting information to diverse extra-linguistic systems: vision models, physics engines, formal logic provers, theory of mind engines, memory augmentations, etc.
12/n

Implications:
1️⃣ Our proposal reconciles the sometimes-broad activation of many brain regions in response to language with decades of neuroimaging studies demonstrating the functional specificity of these same regions.
11/n

Clarification: Our hypothesis is distinct from the claim that the whole brain—not a single, specialized network—is responsible for language processing. It is important to distinguish LANG from the exportation targets, which can be strongly engaged by non-linguistic inputs.
10/n

It remains an open and important question how all of the above information is integrated into a single coherent representation of the situation that the language describes.
9/n

For other brain regions, the empirical record is less clear.
Other possible targets of exportation include the physics network, the navigation system, regions of the ventral pathway, motor regions, amodal semantic regions, and the default network.
8/n

The strongest evidence for exportation comes from brain regions that build mental models of other minds—the rTPJ the rest of the Theory of Mind network. These regions are engaged by language that describes mental content.
7/n

If the core language system does not fully process the meaning of a linguistic input, what other systems do? Hypothesis: Regions that build mental models of minds, objects, and places, that store our memories and word knowledge, and that contain our perceptual and motor reps.!
6/n

This rep. is abstract (no longer tied to particular words), but critically, it is an approximation of meaning, derived solely from our knowledge of lang. statistics. Thus, LANG is only capable of *shallow* understanding, and is not the end point of linguistic processing.
5/n

During language comprehension, LANG’s goal is to extract info from word sequences by recognizing familiar words and figuring out how they go together. The resulting representation captures both structural and co-occurrence-based information encoded in a sentence.
4/n

Cognitive neuroscience may offer important insights! Neuroscientists have leveraged fMRI and related methods to identify and characterize both the *core language system* (LANG) and many other cognitive and neural systems that it may export information to.
3/n

Language understanding entails not just extracting the surface-level meaning of an input, but constructing rich mental models of the _situation_ it describes. Yet, the nature of this rep., and the conditions under which it is constructed, has remained elusive.
2/n