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 tested several language models using 40
classic false-belief tasks widely used to test ToM in humans. The models
published before 2020 showed virtually no ability to solve ToM tasks. Yet, the
first version of GPT-3 ("davinci-001"), published in May 2020, solved about 40%
of false-belief tasks-performance comparable with 3.5-year-old children. Its
second version ("davinci-002"; January 2022) solved 70% of false-belief tasks,
performance comparable with six-year-olds. Its most recent version, GPT-3.5
("davinci-003"; November 2022), solved 90% of false-belief tasks, at the level
of seven-year-olds. GPT-4 published in March 2023 solved nearly all the tasks
(95%). 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.

通过对多个语言模型进行人类 “心灵理论” 测试，研究发现自 2020 年发布的 GPT-3 版本以来语言模型在解决伪信念问题上的表现已经逐渐逼近人类的表现水平

语言模型中的心智理论可能已自发涌现

Theory of Mind May Have Spontaneously Emerged in Large Language Models

Social reasoning necessitates the capacity of theory of mind (ToM), the
ability to contextualise and attribute mental states to others without having
access to their internal cognitive structure. Recent machine learning
approaches to ToM have demonstrated that we can train the observer to read the
past and present behaviours of other agents and infer their beliefs (including
false beliefs about things that no longer exist), goals, intentions and future
actions. The challenges arise when the behavioural space is complex, demanding
skilful space navigation for rapidly changing contexts for an extended period.
We tackle the challenges by equipping the observer with novel neural memory
mechanisms to encode, and hierarchical attention to selectively retrieve
information about others. The memories allow rapid, selective querying of
distal related past behaviours of others to deliberatively reason about their
current mental state, beliefs and future behaviours. This results in ToMMY, a
theory of mind model that learns to reason while making little assumptions
about the underlying mental processes. We also construct a new suite of
experiments to demonstrate that memories facilitate the learning process and
achieve better theory of mind performance, especially for high-demand
false-belief tasks that require inferring through multiple steps of changes.

该研究采用新颖的神经记忆机制和分层注意力相结合的方法设计了一个理论心智模型，ToMMY，使其能够快速准确地推断他人的意图、信念与将来的行为，进一步实验证明神经记忆机制可在高难度的虚假信念任务中提高心理理解的准确性。