Epistemological challenges in the study of “Theory of Mind” in LLMs and humans

Value of the Information & Strength of the Argument

The talk provides valuable insights into the use of LLMs as model organisms for studying human cognition, particularly Theory of Mind. Trott presents empirical evidence from his own studies, showing that LLMs exhibit some sensitivity to belief states but fall short of human performance. He carefully discusses limitations, such as data contamination and reproducibility, and proposes a nuanced framework (differential construct validity) for interpreting LLM performance. The argumentation is solid, building from specific experiments to broader epistemological considerations, and he acknowledges the work-in-progress nature of his theoretical contributions.

Scientific Rigor, Source Quality, Title Accuracy

Trott demonstrates scientific rigor by referencing his published papers (Jones et al., 2024; Trott et al., 2023) and discussing methodological details. He addresses potential criticisms of the false belief task and the challenges of studying closed-source models. The title accurately reflects the content, focusing on epistemological challenges. The talk is well-structured and grounded in empirical evidence, though some parts are based on unpublished work, which is clearly indicated.

The title accurately reflects the content, which focuses on epistemological challenges in studying Theory of Mind in LLMs and humans.

8/10

The talk is given by a researcher with relevant expertise, presents empirical studies with published results, and discusses methodological limitations. However, it is a seminar presentation, not a peer-reviewed article, and some claims are based on unpublished work.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction to the talk and the concept of LLM-ology.
• 05:00Discussion of Theory of Mind and the language exposure hypothesis.
• 10:00Explanation of using LLMs as model organisms and the false belief task.
• 15:00Description of the 2023 study with GPT-3 and the results.
• 20:00Replication with 55 open-weight models and findings.
• 25:00Discussion of three interpretive views: duck test, axiomatic rejection, differential construct validity.
• 30:00Theoretical perspectives: plausibility and face validity.
• 35:00Empirical perspectives: convergent validity and process mechanisms.
• 40:00Conclusion and call for methodological innovation.

• Comparing humans and large language models on an Experimental Protocol Inventory for Theory of Mind Evaluation (EPITOME) — Referenced as a publication by Jones, Trott, & Bergen (2024) in the talk.
• Do large language models know what humans know? — Referenced as a publication by Trott, Jones, Chang, Michaelov, & Bergen (2023) in the talk.

• Theory of Mind — General reference for the concept.
• False belief task — Background on the task used in the studies.