The Emergence of ToM in LLMs

This has already been posted but I'm afraid the title may have kept it low in the SEO race that is hacker news.

Of note:

    "Although the results from vanilla prompts and the user study build the illusion of ToM abilities in LLMs, we conclude that our perturbation experiments are sufficient to clarify that ToM abilities are absent from LLMs. ToM evaluation must entail evaluation of reasoning abilities with minimal leniency towards failure cases."

It's worth noting that the word "idempotent" does not occur anywhere in the text.

It seems like a latent conversation given how ToM has impacted everything from Autism research to consciousness research.

I may have missed the deeper discussion in my cursory reading, which is lossy and poor. If so, I'm sorry.

1) The idea that there are extant bidirectional games, whether intentional or unintentional, is absent, whether elicited by the LLM or the human interlocutor in conversation. These are best labeled as mentalism: the creation of illusions as a technique to delude the human or bypass the conditional guardrails of the LLM.

2) The introduction of bicameralism (Jaynes) along with ideas stemming from the "controlled hallucination" theory of Anil Seth and its further explication by Andy Clark is also absent.

I'd be happy if anyone has insight into these areas beyond Peter Pirolli's work:

https://www.efsa.europa.eu/sites/default/files/event/180918-...

on Information Foraging and so on.