TL;DR
Do large language models (LLMs) truly understand the social world through unified internal representations, or are they just memorizing disjointed conversational tricks? This paper applies cognitive neuroscience techniques to show that LLMs recruit shared, integrated computational mechanisms for both general Theory of Mind (ToM) and pragmatic language reasoning. By locating and ablating these "social subnetworks," the authors demonstrate that LLMs develop emergent, interconnected social world models rather than isolated competencies.
What Problem Does This Paper Address?
While LLMs show impressive capabilities, debate continues over whether they are merely "stochastic parrots" or possess genuine internal "world models." While world models have been successfully demonstrated in narrow, rule-bound environments (like Chess or Othello), it is less clear if they exist for complex, ambiguous human domains like social interaction.
Specifically, this paper investigates the relationship between Theory of Mind (ToM)—the ability to attribute unobservable mental states like beliefs and desires to others—and pragmatic reasoning (understanding context-dependent meaning, such as irony, hints, or sarcasm). The authors tackle a fundamental question: Do LLMs use separate, specialized circuits for these tasks (the functional specialization hypothesis), or do they share a unified "social world model" (the functional integration hypothesis), much like the overlapping neural networks found in the human brain?
Methods
The authors use a two-pronged approach combining behavioral and mechanistic analysis:
- Behavioral Evaluation: They test 48 different LLMs across dozens of tasks to see if ToM and pragmatic performance correlate and predict each other better than general language proficiency.
- Functional Localization & Ablation: Borrowing a technique from cognitive neuroscience, they identify specific "subnetworks" (groups of neurons) within the LLMs that are highly active during ToM tasks compared to closely matched control tasks (e.g., reading a story about a character's false belief vs. a story about a physically outdated photograph). They then perform causal ablation experiments—setting the activations of these localized neurons to zero—to observe the impact on the model's performance across different domains.
Datasets
The study leverages a massive evaluation suite consisting of:
- Pragmatics Benchmarks: 16 datasets covering 10 phenomena (e.g., scalar implicatures, irony, sarcasm, indirect speech) drawn from benchmarks like Ma et al. (2025) and PUB.
- Theory of Mind (ToM) Benchmarks: 22 datasets categorized under the ATOMS framework (which breaks ToM down into beliefs, intentions, desires, emotions, knowledge, percepts, and non-literal communication), including BigToM, OpenToM, and ToMi.
- Linguistic Baselines: BLiMP (grammar) and SNLI (natural language inference).
- Novel Localizer Suite: The authors introduced a large, synthetically generated dataset (via GPT-5) containing 1,400 stimuli across four suites (LatentBeliefs, CommunicativeIntent, GameBeliefs, MoralIntent) based on validated human fMRI experimental materials.
Key Insights and Ideas
- Social Integration over Specialization: LLMs do not treat social reasoning and conversational pragmatics as completely separate tasks. Instead, they cross-purpose internal mechanisms, suggesting that efficient compression principles during pretraining favor the emergence of a unified social world model.
- Causal Proof of Social Circuits: The most striking "aha" moment comes from the ablation studies. Damaging the specific neurons localized for ToM didn't just hurt ToM performance; it systematically degraded pragmatic reasoning as well, while leaving general grammatical and linguistic abilities largely intact.
- Shared Tracking Mechanisms: Further exploration revealed that these social subnetworks are also heavily involved in tracking physical entities (e.g., box contents over a series of shifts), suggesting that social reasoning in LLMs builds directly upon foundational tracking mechanisms.
Key Results
- Across 48 models, there is a strong, significant positive correlation ($r = 0.68$) between average accuracy on ToM and pragmatics tasks.
- Statistical modeling showed that a model's ToM performance is a significantly better predictor of its pragmatic accuracy than its performance on general linguistic benchmarks (BLiMP and SNLI).
- In the global ablation analysis, knocking out the localized ToM subnetworks caused a significant drop in both ToM and pragmatic performance, but had no credible negative effect on baseline linguistic benchmarks. This asymmetry strongly supports the functional integration hypothesis.
Limitations and Future Work
- Coarse Interventions: Setting neuron activations to zero is a relatively blunt instrument. The authors note that future work should employ finer-grained mechanistic interpretability tools like activation patching or circuit discovery.
- English and Cultural Bias: All datasets were exclusively in English. Pragmatics and social norms vary wildly across cultures, meaning these findings might not seamlessly generalize to other languages.
- Lack of Human Validation for Stimuli: While the synthetic localizer suite was inspired by human fMRI studies, the newly generated text prompts have not been validated on actual human brains to confirm they activate the same biological pathways.
- Data Contamination Risk: Because some benchmarks have been public for years, there is an inherent risk that newer models encountered them during pretraining, potentially inflating baseline scores.
Summary updated