On Emergent Social World Models -- Evidence for Functional Integration of Theory of Mind and Pragmatic Reasoning in Language Models

TL;DR

This study interrogates whether large language models develop emergent social world models by testing whether Theory of Mind (ToM) reasoning and pragmatic language use share computational resources. It combines a broad behavioral evaluation (16 pragmatics datasets and 22 ToM benchmarks across 48 models) with cognitive neuroscience–inspired functional localization (eight subnetworks across four localizer suites, plus ablation tests) grounded in the ATOMS framework. Results reveal a consistent, though modest, link between ToM and pragmatics performance, with ToM accuracy providing predictive gains for pragmatics and ATOMS facets such as intentions, desires, emotions, percepts, and non-literal communication driving these gains; ablations implicate shared subnetworks as causally involved in both domains, supporting a functional integration view. The work advances the field by providing novel ToM localizer data, refining localization methods, and offering nuanced insights into how social cognition may be realized in AI systems, while noting limitations and the need for replication and cross-linguistic validation.

Abstract

This paper investigates whether LMs recruit shared computational mechanisms for general Theory of Mind (ToM) and language-specific pragmatic reasoning in order to contribute to the general question of whether LMs may be said to have emergent "social world models", i.e., representations of mental states that are repurposed across tasks (the functional integration hypothesis). Using behavioral evaluations and causal-mechanistic experiments via functional localization methods inspired by cognitive neuroscience, we analyze LMs' performance across seven subcategories of ToM abilities (Beaudoin et al., 2020) on a substantially larger localizer dataset than used in prior like-minded work. Results from stringent hypothesis-driven statistical testing offer suggestive evidence for the functional integration hypothesis, indicating that LMs may develop interconnected "social world models" rather than isolated competencies. This work contributes novel ToM localizer data, methodological refinements to functional localization techniques, and empirical insights into the emergence of social cognition in artificial systems.

Figures (13)

• Figure 1: Average accuracy of 48 models on ToM (x-axis) and pragmatics (y-axis) tasks.
• Figure 2: Percentage of localized units across model layers (x-axis) identified by different target subnetwork localizers (colors) for Qwen-2.5 Instruct models (shapes).
• Figure 3: All: Plots show the mean causal effects (y-axis), i.e., difference in accuracy relative to the non-ablated baseline when ablating the critical (ToM) subnetworks or the control (least active) sub-networks (x-axis), for different test sets (colors and shapes). 95% CIs indicate change in by-dataset accuracy across localizer suites. Top row: Fictitious, stylized result that could come from our experiments (left-most plot), and a visualization of the three main predictions made by the functional integration hypothesis. Bottom row: Four observed outcomes at different levels of analysis. Results from the "global analysis" (marginalizing over all language models and localizers) are shown on the left; followed by example results for data from only one language model (averaging over localizers) and two localizer sets (averaging over language models). Predictions are supported for the "global analysis," but not for all subunits at finer levels of analysis (e.g., the right-most example in the bottom row).
• Figure 4: Average accuracy (y-axis) for the domains pragmatics, theory of mind and syntax (color) by model size (x-axis).
• Figure 5: Detailed overview of the accuracy (y-axis) results of all models (x-axis) on the different pragmatic datasets (facets).