RuCCS - Dr. Sean Trott - In Person Talk - Tuesday, February 10, 2026

Описание: Humans regularly reason about the belief states of others. Where does this capacity originate? Competing hypotheses include biologically evolved endowments, social interaction, and exposure to language. Recent advances in language models (LMs) offer a novel opportunity to test whether sensitivity to belief states can emerge in principle from exposure to the distributional statistics of language alone. In this talk, I present evidence testing this hypothesis using a sample of 41 open-weight LMs: I find that larger LMs are indeed sensitive to the implied belief states of characters in written passages—but that this sensitivity reliably falls short of most humans tested. These findings suggest that while distributional statistics may be sufficient to account for some sensitivity to mental states, they are (currently) insufficient to account for human-level behavior on these tasks. I then turn to critical epistemological challenges that emerge when using LMs as “model organisms”. Does the same experimental task measure the same underlying construct in humans and LMs, or does it suffer from “differential construct validity”? I explore empirical and theoretical approaches to these challenges, including convergent validity analyses, developmental comparisons, and identifying the mechanistic underpinnings of model behavior. I conclude by arguing that these epistemological challenges represent a crucial opportunity for methodological innovation and theoretical refinement in both Cognitive Science and AI research.