Skip to content
The methodOpen data
Back to the collection

Evaluating Large Language Models in Theory of Mind Tasks

Michal Kosinski

2023 · arXiv

Language models
Read the paper
HN score
0
Mentions
10
Readers
9
Months
4
What’s in a score?

From the abstract

Eleven Large Language Models (LLMs) were assessed using a custom-made battery of false-belief tasks, considered a gold standard in testing Theory of Mind (ToM) in humans. The battery included 640 prompts spread across 40 diverse tasks, each one including a false-belief scenario, three closely matched true-belief control scenarios, and the reversed versions of all four. To solve a single task, a model needed to correctly answer 16 prompts across all eight scenarios. Smaller and older models solved no tasks; GPT-3-davinci-003 (from November 2022) and ChatGPT-3.5-turbo (from March 2023) solved 20% of the tasks; ChatGPT-4 (from June 2023) solved 75% of the tasks, matching the performance of six-year-old children observed in past studies. We explore the potential interpretation of these findings, including the intriguing possibility that ToM, previously considered exclusive to humans, may have spontaneously emerged as a byproduct of LLMs' improving language skills.

Metadata: arXiv. Abstracts retain their original authors’ rights.

In the conversation

Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels

10 selected excerpts · 10 total mentions

netfortius
Neutral mention · rule
One of the articles linked here [1] reminded me of the great Ranulph Glanville, and his "Inside Every White Box There Are Two Black Boxes Trying To Get Out" [1] https://arxiv.org/abs/2302.02083
Original comment This time, it feels different
slowmovintarget
Neutral mention · rule
Another poster shared a link to this paper last week for Theory of Mind: https://arxiv.org/abs/2302.02083 We're seeing those other capabilities emerge; like being able to play chess though it's not been trained to do so. That is, these LLMs are displaying emergent abilities associated with reasoning. These LLMs aren't R. Daneel Olivaw or R2D2 (which is what I think of when I think of the original term for AI, and what we took to calling AGI). We're closer to seeing the just-the-facts AIs we encounter in Blindsight. Intelligence without awareness. Funny that we still have to use science fiction to make our comparisons because our philosophy of intelligence, mind, and consciousness are insufficient to speak on the matter clearly.
Original comment Parent story unavailable in snapshot
agravier
Neutral mention · rule
It matters to understand how things work in order to understand their behaviour and react properly. I've seen people draw conclusions from applying Theory of Mind tests to LLMs (https://arxiv.org/abs/2302.02083). Those psychological were designed to assess humans psychological abilities or deficiencies, they assume that the language used by the human respondent reflects their continued and deep understanding of others' state of mind. In LLMs, there is no understanding involved. Dismissing the Chinese Room argument is an ostrich strategy. You're refusing to consider its lack of understanding despite knowing pretty well how an LLM work, because you don't want to ascribe understanding to humans, I suppose?
Original comment Bing: “I will not harm you unless you harm me first”
ec109685
Neutral mention · rule
There's more than comprehension. It can do some amount of higher order reasoning: https://arxiv.org/abs/2302.02083 "Theory of mind (ToM), or the ability to impute unobservable mental states to others, is central to human social interactions, communication, empathy, self-consciousness, and morality. We administer classic false-belief tasks, widely used to test ToM in humans, to several language models, without any examples or pre-training. Our results show that models published before 2022 show virtually no ability to solve ToM tasks. Yet, the January 2022 version of GPT-3 (davinci-002) solved 70% of ToM tasks, a performance comparable with that of seven-year-old children.…
Read the full comment
Moreover, its November 2022 version (davinci-003), solved 93% of ToM tasks, a performance comparable with that of nine-year-old children. These findings suggest that ToM-like ability (thus far considered to be uniquely human) may have spontaneously emerged as a byproduct of language models' improving language skills."
Original comment Bing AI Can't Be Trusted
sdenton4
Neutral mention · rule
"Theory of Mind May Have Spontaneously Emerged in Large Language Models" https://arxiv.org/abs/2302.02083 'Theory of mind (ToM), or the ability to impute unobservable mental states to others, is central to human social interactions, communication, empathy, self-consciousness, and morality. We administer classic false-belief tasks, widely used to test ToM in humans, to several language models, without any examples or pre-training. Our results show that models published before 2022 show virtually no ability to solve ToM tasks. Yet, the January 2022 version of GPT-3 (davinci-002) solved 70% of ToM tasks, a performance comparable with that of seven-year-old children.…
Read the full comment
Moreover, its November 2022 version (davinci-003), solved 93% of ToM tasks, a performance comparable with that of nine-year-old children. These findings suggest that ToM-like ability (thus far considered to be uniquely human) may have spontaneously emerged as a byproduct of language models' improving language skills.'
Original comment I had ChatGPT write a Decemberists song
jointpdf
Neutral mention · rule
I see a lot of confident assertions of this type (LLMs don’t actually understand anything, cannot be creative, cannot be conscious, etc.), but never any data to substantiate the claim. This recent paper suggests that recent LLMs may be acquiring theory of mind (or something analogous to it): https://arxiv.org/abs/2302.02083v1 Some excerpts: > We administer classic false-belief tasks, widely used to test ToM in humans, to several language models, without any examples or pre-training. Our results show that models published before 2022 show virtually no ability to solve ToM tasks. Yet, the January 2022 version of GPT- 3 (davinci-002) solved 70% of ToM tasks, a performance…
Read the full comment
comparable with that of seven-year-old children. Moreover, its November 2022 version (davinci-003), solved 93% of ToM tasks, a performance comparable with that of nine-year-old children. > Large language models are likely candidates to spontaneously develop ToM. Human language is replete with descriptions of mental states and protagonists holding divergent beliefs, thoughts, and desires. Thus, a model trained to generate and interpret human-like language would greatly benefit from possessing ToM. > While such results should be interpreted with caution, they suggest that the recently published language models possess the ability to impute unobservable mental states to others, or ToM. Moreover, models’ performance clearly grows with their complexity and publication date, and there is no reason to assume that their it should plateau anytime soon. Finally, there is neither an indication that ToM-like ability was deliberately engineered into these models, nor research demonstrating that scientists know how to achieve that. Thus, we hypothesize that ToM-like ability emerged spontaneously and autonomously, as a byproduct of models’ increasing language ability.
Original comment I had ChatGPT write a Decemberists song

Keep following the thread.

More in Language models

More from the same topics.