Like there are plenty of shortcomings of LLMs but it feels like people are comparing them to some platonic ideal human when writing them off
ToM is a large topic, but most people, when talking about an entity X, they have a state in memory about that entity, almost like an Object in a programming language. Thta Object has attributes, and conditions etc that exist beyond the context window of the observer.
If you have a friend Steve, who is a doctor. And you don't see him for 5 years, you can predict he will still be working at the hospital, because you have an understanding of what Steve is.
For an LLM you can define a concept of Steve, and his profession and it will adequately mimic replies about him. But in 5 years that LLMs would not be able to talk about Steve. It would recreate a different conversation, possibly even a convincing simulacrum of remembering Steve. But internally, there is no Steve, nowhere in the nodes of the LLM does Steve exist or have ever existed.
That inability to have a world model means that an LLM can replicate the results of a theory of mind but not posses one.
Humans lose track of information, but we have a state to keep track of elements that are ontologicaly distinct. LLMs do not, and treat them as equal.
For a human, the sentence Alice and bob go to the market, when will they be back? is different than Bob and Alice went to the market, when will they be back?
Because Alice and Bob are real humans, you can imagine them, you might have even met them. But to an LLM those are the same sentence. Even outside of the argument about The Red Room/ Mary's room there simply are enough gaps in the way a LLM is constructed to be considered a valid owner of a ToM
I don't think we have any strong evidence on whether LLMs have world-models one way or another - it feels like a bit of a fuzzy concept and I'm not sure what experiments you'd try here.
I disagree with your last point, I think those are functionally the same sentence
In that sentence you are implying that you have the "ability to model ... another". An LLM cannot do that, it can't have an internal model that is consistent beyond its conversational scope. Its not meant to. Its a statistics guesser, its probabilistic, holds no model, and its anthropomorphised by our brains because the output is incredibly realistic not because it actually has that ability
The ability to mimic the replies of someone with that ability, is the same of Mary being able to describe all the qualities of Red. She still cannot see red, despite her ability to pass any question in relation to its characteristics.
> I don't think we have any strong evidence on whether LLMs have world-models one way or another
They simply cannot by their architecture. Its a statistical language sampler, anything beyond the scope of that fails. Local coherance is why they pick the next right token not because they can actually model anything.
> I think those are functionally the same sentence
Functionally and literally are not the same thing though. Its why we can run studies as to why some people might say Bob and Alice (putting the man first) or Alice and Bob (alphabetical naming) and what human societies and biases affect the order we put them on.
You could not run that study on an LLM because you will find that statistically speaking the ordering will be almost identical to the training data. If the training data overwhelmingly puts male names first or whether the training data orders list alphabetically you will see that reproduced on the output of the llm because Bob and Alice are not people, they are statistical probably letters in order.
LLM seem to trigger borderline mysticism in people who are otherwise insanely smart, but the kind of "we cant know its internal mind" sounds like reading tea leaves, or horoscopes by people with enough Phds to have their number retired on their university like Michael Jordan.
If you can rigorously state what "having a world model" consists of and what - exactly - about a transformer architecture precludes it from having one I'd be all ears. As would the academic community, it'd be a groundbreaking paper.
You dont need to be unbelievably confident or understand exactly how AI and human brains work to make certain assesments. I have a limited understanding of biology, I can however make an assesment on who is healthier between a 20 year old person who is active and has a healthy diet compared to someone with a sedentary lifestyle, in their late 90s and with a poor diet. This is an assesement we can do despite the massive gaps we have in terms of understanding aging, diet, activity and overall health impact of individual actions.
Similarly, despite my limited understanding of space flight, I know Apollo 13 cannot cook an egg or recite french poetry. Despite the unfathamobly cool science inside the space craft, it cannot, by design do those things.
> the field of mechanistic interpretability
The field is cool, but it cannot prove its own assumption yet. The field is trying to prove you can reverse engineer a model to be humanly understood. Their assumptions such as mapping specific weights or neurons to features has failed to be reproduced multiple times, with the weight effects being way more distributed and complicated than initially thought. This is specially true for things that are equally mystified as the emergent abilities of LLMs. The ability of mimicking nuanced language being unlocked after a critical mass of parameters, does not create a rule as for which increased parameterisation will increase linerly or exponentially the abilities of an LLM.
> it turns out, to predict text really really well, building an internal model of the underlying distribution works really well
yeah, an internal model works well because most words are related to their neighbours, thats the kind of local coherance the model excels at. But to build a world model, the kind a human mind interacts with, you need a few features that remain elusive (some might argue impossible to achieve) to a transformer architecture.
Think of games like chess, an llm is capable of accurately expressing responses that sound like game moves, but the second the game falls outside its context window the moves become incoherent (while still sounding plausible).
You can fix this, with arch that do not have a transformer model underlying it, or by having multiple agents performing different tasks inside your arch, or by "cheating" and using a state outside the llm response to keep track of context beyond reasonable windows. Those are "solutions" but all just kinda prove the transformer lacks that ability.
Other tests abour casuality, or reacting to novel data (robustness), multi step processes and counterfactual reasoning are all the kind of tasks transformers still (and probably always) will have trouble with.
For a tech that is so "transparent" in its mistakes, and so "simple" in its design (replacing the convolutions with an attention transformer, its genius) I still think its talked about in borderline mystic tones, invoking philosophy and theology, and a hope for AGI that the tech itself does not lend to beyond the fast growth and surprisingly good results with little prompt engineering.