This highlights one of the types of muddled thinking around LLMs. These tasks are used to test theory of mind because for people, language is a reliable representation of what type of thoughts are going on in the person's mind. In the case of an LLM the language generated doesn't have the same relationship to reality as it does for a person.
What is being demonstrated in the article is that given billions of tokens of human-written training data, a statistical model can generate text that satisfies some of our expectations of how a person would respond to this task. Essentially we have enough parameters to capture from existing writing that statistically, the most likely word following "she looked in the bag labelled (X), and saw that it was full of (NOT X). She felt " is "surprised" or "confused" or some other word that is commonly embedded alongside contradictions.
What this article is not showing (but either irresponsibly or naively suggests) is that the LLM knows what a bag is, what a person is, what popcorn and chocolate are, and can then put itself in the shoes of someone experiencing this situation, and finally communicate its own theory of what is going on in that person's mind. That is just not in evidence.
The discussion is also muddled, saying that if structural properties of language create the ability to solve these tasks, then the tasks are either useless for studying humans, or suggest that humans can solve these tasks without ToM. The alternative explanation is of course that humans are known to be not-great at statistical next-word guesses (see Family Feud for examples), but are also known to use language to accurately describe their internal mental states. So the tasks remain useful and accurate in testing ToM in people because people can't perform statistical regressions over billion-token sets and therefore must generate their thoughts the old fashioned way.
What is being demonstrated in the article is that given billions of tokens of human-written training data, a statistical model can generate text that satisfies some of our expectations of how a person would respond to this task. Essentially we have enough parameters to capture from existing writing that statistically, the most likely word following "she looked in the bag labelled (X), and saw that it was full of (NOT X). She felt " is "surprised" or "confused" or some other word that is commonly embedded alongside contradictions.
What this article is not showing (but either irresponsibly or naively suggests) is that the LLM knows what a bag is, what a person is, what popcorn and chocolate are, and can then put itself in the shoes of someone experiencing this situation, and finally communicate its own theory of what is going on in that person's mind. That is just not in evidence.
The discussion is also muddled, saying that if structural properties of language create the ability to solve these tasks, then the tasks are either useless for studying humans, or suggest that humans can solve these tasks without ToM. The alternative explanation is of course that humans are known to be not-great at statistical next-word guesses (see Family Feud for examples), but are also known to use language to accurately describe their internal mental states. So the tasks remain useful and accurate in testing ToM in people because people can't perform statistical regressions over billion-token sets and therefore must generate their thoughts the old fashioned way.