You're right of course. Let me rephrase. What I was trying to say was: If the model solves some reasoning tasks that may imply it's building some inner world model which is fundamental to reasoning. That in-turn might mean that the model is reasoning.

Others have argued against the "inner world model" theory and suggested that solving reasoning tasks is merely an extension of the "stochastic parrot" scenario - i.e., they claim that no such world model exists and that the model has rote memorized these reasoning scenarios.

Why not just give it a series of directions e.g. go one step forwards, turn right, etc like a LOGO program, and then ask it if it is back to the start or not. If the series of instructions is randomly generated then this is impossible to solve without a world model.

People have tried this! And for not-complex problems, the newer models do give the right answer and yet usually fail on the complex tasks.

Prompting does help though - if you ask the model to check it's own answer, it can often catch errors it made and improve. But the interesting part is precisely how does the model succeed at all at any spatial reasoning task? Surely it hasn't seen all spatial reasoning tasks in its training set. So, what's the other hypothesis? As you say, people suggest that there could be a world model that is constructed internally by the LLM.

However, as you can see right here in this thread that people disagree that you need a world model for solving such tasks.

From the preface to the 20th anniversary edition of Godel, Escher, Bach:

"Meaning cannot be kept out of formal systems when sufficiently complex isomorphisms arise. Meaning comes in despite one's best efforts to keep symbols meaningless! ...When a system of "meaningless" symbols has patterns in it that accurately track, or mirror, various phenomena in the world, then that tracking or mirroring imbues the symbols with some degree of meaning -- indeed, such tracking or mirroring is no less and no more than what meaning is. Depending on how complex and subtle and reliable the tracking is, different degrees of meaningfulness arise."

In other words, when one can reliably ask a language model a question and get a sensible answer, one is forced to conclude that it does in some sense "understand" what it is saying. This is also I think the essential philosophical thrust of the Turing Test, which is often misunderstood as a mere benchmark.

(I notice a common objection to examples of LLMs clearly demonstrating understanding is "it saw something similar in the training set". That may be true (though unfalsifiable) in any given instance, but the number of permutations of things LLMs correctly respond to far exceed the size of any training set. They are certainly generalizing, and interpreting their inputs on a conceptual level.)

It’s a type of understanding, but one we are not really familiar with, because it seems to understand tasks that it was trained on, but fails and other very basic ways on simple but different tasks.