Social World Models

TL;DR

This paper introduces social world models (SWMs) and a novel structured representation (S3AP) to improve AI's understanding of implicit social dynamics, significantly enhancing social reasoning and interaction capabilities.

Contribution

The paper proposes SWMs and S3AP formalism to explicitly model unobserved social states, improving AI social reasoning over traditional methods.

Findings

+51% improvement on FANToM benchmark

Up to +18% improvement on SOTOPIA benchmark

Explicit modeling of mental states is more effective than baseline methods

Abstract

Humans intuitively navigate social interactions by simulating unspoken dynamics and reasoning about others' perspectives, even with limited information. In contrast, AI systems struggle to structure and reason about implicit social contexts, as they lack explicit representations for unobserved dynamics such as intentions, beliefs, and evolving social states. In this paper, we introduce the concept of social world models (SWMs) to characterize the complex social dynamics. To operationalize SWMs, we introduce a novel structured social world representation formalism (S3AP), which captures the evolving states, actions, and mental states of agents, addressing the lack of explicit structure in traditional free-text-based inputs. Through comprehensive experiments across five social reasoning benchmarks, we show that S3AP significantly enhances LLM performance-achieving a +51% improvement on FANToM over OpenAI's o1. Our ablations further reveal that these gains are driven by the explicit modeling of hidden mental states, which proves more effective than a wide range of baseline methods. Finally, we introduce an algorithm for social world models using S3AP, which enables AI agents to build models of their interlocutors and predict their next actions and mental states. Empirically, S3AP-enabled social world models yield up to +18% improvement on the SOTOPIA multi-turn social interaction benchmark. Our findings highlight the promise of S3AP as a powerful, general-purpose representation for social world states, enabling the development of more socially-aware systems that better navigate social interactions.

Tables

Table 1. Table 1: Performance comparison of models with CoT and with S 3 AP representations (generated by O3) on various social reasoning tasks. Bolded values indicate the best average performance across models. L4 means Llama 4, 4o means GPT-4o, o3m means o3-mini.

Model	ToMi	Para	HiToM	FAN	MMToM
L4 CoT	0.655	0.740	0.720	0.264	0.443
L4 S3AP	0.662	0.763	0.700	0.415	0.450
4o CoT	0.813	0.818	0.660	0.396	0.652
4o S3AP	0.927	0.905	0.750	0.491	0.692
o3m CoT	0.863	0.817	0.810	0.057	0.493
o3m S3AP	0.960	0.900	0.860	0.170	0.506
R1 CoT	0.945	0.893	0.420	0.491	0.374
R1 S3AP	0.980	0.950	0.510	0.547	0.437
o1 CoT	0.952	0.835	0.870	0.415	0.725
o1 S3AP	0.985	0.932	0.880	0.623	0.785

Table 2. Table 2: ParaToMi accuracy (%) with CoT vs S 3 AP on Qwen3 models. Δ \Delta (pp) stands for the absolute performance difference .

Model	CoT	S³AP	$\Delta$ (pp)
Qwen3-80B	77	91	+14
Qwen3-32B	75	88	+13
Qwen3-7B	44	45	+1
Qwen3-0.6B	48	50	+2

Table 3. Table 3 : Performance comparison on ConfAIde benchmark with and without S 3 AP representations.

Model	ConfAIde	ConfAIde w S3AP
GPT-4o	0.575	0.740
GPT-4.1	0.995	0.985
o1	0.980	0.975
o3	0.910	0.955
o3-mini	0.765	0.770
R1	0.835	0.785
Llama 4	0.810	0.820
Average	0.839	0.861

Table 4. Table 4 : Performance comparison of models with and without S 3 AP representations on various social reasoning tasks with o1 -powered S 3 AP-Parser.

Model	ParaToMi	HiToM	MMToM-QA	FANToM	Confaide
GPT-4o	0.818	0.660	0.652	0.396	0.575
GPT-4o w S3AP	0.885	0.750	0.692	0.472	0.625
GPT-4.1	0.802	0.830	0.586	0.528	0.995
GPT-4.1 w S3AP	0.892	0.760	0.583	0.623	1.000
o1	0.835	0.870	0.725	0.415	0.980
o1 w S3AP	0.933	0.880	0.761	0.528	0.990
o3	0.955	0.930	0.715	0.547	0.910
o3 w S3AP	0.947	0.890	0.722	0.698	0.975
o3-mini	0.817	0.810	0.493	0.057	0.780
o3-mini w S3AP	0.863	0.790	0.460	0.151	0.750
R1	0.893	0.420	0.374	0.491	0.835
R1 w S3AP	0.932	0.520	0.412	0.585	0.840
Llama 4	0.740	0.720	0.443	0.264	0.810
Llama 4 w S3AP	0.745	0.680	0.453	0.321	0.785
AVG	0.837	0.749	0.570	0.385	0.841
AVG w S3AP	0.885	0.753	0.583	0.483	0.852

Equations

$$p({\mathcal{A}}_t^{-i}\mid {\mathcal{S}}_t)$$

$$p({\mathcal{S}}_{t+1}\mid {\mathcal{S}}_t,{\mathcal{A}}_t^{-i},a_t^i)$$

$$\text{overall\_score}=0.3\times \text{structural\_score}+0.7\times \text{observation\_accuracy}.$$

Full text

Machine Learning, ICML

1 Introduction

Unlocking social intelligence is an elusive yet foundational challenge of AI (Gunning, 2018). For AI systems to effectively interact with humans, they must be able to both understand and model complex social dynamics, requiring reasoning about others’ mental states, tracking how beliefs evolve, and interpreting perspectives within social contexts (Sap et al., 2023; Tomasello, 2009). However, despite rapid progress in general-purpose reasoning capabilities, current AI systems still lack the core mechanisms needed for mentalizing and navigating social contexts (Shapira et al., 2023; Yerukola et al., 2024; Kim et al., 2023b).

This limitation stems from two fundamental challenges: 1) AI systems primarily learn social dynamics from static texts (Sap et al., 2023), descriptions of situations, and narratives. These input representations are inherently lossy and suffer from reporting biases: mention only salient events (Gordon & Van Durme, 2013), omit explicit mentions of mental states and perspectives (Lucy & Gauthier, 2017), and often present an all-knowing viewpoint that fails to capture the partial, subjective nature of real social experiences (Fischbach et al., 2021; Epstein, 1999; Mar & Oatley, 2008; Mani, 2012). 2) Humans routinely construct rich internal models to interpret partial and biased inputs (Frith & Frith, 2006; Johnson-Laird, 1983; Hinsz, 1995), but current AI systems lack computational frameworks designed for recursively reasoning about others’ perspectives and intentions.

We argue that solving these gaps requires the conceptual formulation of Social World Models (SWMs)–computational frameworks that maintain structured representations of social environments, tracking agents’ mental states, beliefs, intentions, and their dynamic interactions over time. Like traditional world models (Ha & Schmidhuber, 2018; Beohar & Melnik, 2022), SWMs capture state transitions, but critically they also encode the social fabric of interaction. Historically, world models have ignored the social dimension because it is difficult to model what is not explicitly present in the input representation: the implicit social dynamics and mental state information that drive human interactions but remain absent from text-based inputs.

To unlock effective SWMs, we introduce S3AP,111Structured Social Simulation Analysis Protocol a novel general-purpose structured social world formalism designed to bridge lossy narrative inputs and the rich representations needed for social reasoning. Specifically, S3AP captures the state of the social world by structuring information extracted from diverse, lossy, and free-form narratives. Following recent generative social simulation systems (Zhou et al., 2024; Hou et al., 2025; Liang et al., 2025; Park et al., 2023), S3AP outlines social agents’ action, perspectives, and environment state at each timestep, reducing ambiguity from free-form text narratives (Figure 2). Inspired by reinforcement learning theories, S3AP connects social reasoning with the rich literature on planning and embodied agents (Ha & Schmidhuber, 2018; Beohar & Melnik, 2022). Designed to be minimal and flexible, this structured representation of the social world state enables seamless integration into LLM-powered generative social simulation systems.

To demonstrate the effectiveness of S3AP as a general-purpose representation of the social world, we develop an LLM-powered S3AP-Parser that automatically converts free-text narratives into structured representations. We show that the parsed structured data enhances LLMs’ performance in reasoning about social interactions, achieving SOTA results across a diverse set of social reasoning tasks including theory of mind reasoning (Sclar et al., 2023), multi-party belief tracking in daily dialogue (Kim et al., 2023b), and embodied social reasoning (Jin et al., 2024). Further ablation studies show that smaller LLMs (e.g., o3-mini) can effectively parse static text into S3AP data, even aiding more capable models at social reasoning (e.g., improving accuracy on ParaToMi of o1 from 83.5% to 94.3%). The consistent improvements across diverse models indicate that our automatically parsed structured representation improves LLMs’ ability to perform social reasoning from a static third-person perspective.

Building upon this structured representation, we then show how to effectively induce and use social world models from S3AP. Inspired by previous works building (non-social) world models (Ha & Schmidhuber, 2018; Xiang et al., 2023), we show that the induced social world model can help AI agents better engage in social interactions. Through experiments on the SOTOPIA platform (Zhou et al., 2024), we demonstrate that agents equipped with a social world model can make more goal-oriented and strategic decisions in social interactions, largely outperforming baseline agents without such world models. Notably, we find that social world models provide distinct advantages in cooperative versus competitive settings. These results highlight that our formalism supports interactive first-person social reasoning, enabling agents to interpret and act more intelligently within social situations from their own perspective.

2 Related Work

Status and Limits of Current LLMs at Social Tasks.

Recent benchmarks show steady progress in LLMs’ social reasoning abilities. False belief tests (Kim et al., 2023b; Wu et al., 2023) evaluate mental state tracking, while social commonsense and norm adherence tasks (Sap et al., 2019; Zhou et al., 2023) probe broader social understanding. These benchmarks adopt an third-person observer setup, where LLMs see the full context and reason from an outsider’s perspective. In contrast, tasks like SOTOPIA (Zhou et al., 2024) and NegotiationArena (Bianchi et al., 2024) embed LLMs as interactive agents navigating social goals from a interactive, first-person view. Both settings have exposed persistent gaps in AI’s ability for long-term theory of mind, and safe social behavior.

Algorithms to Improve LLMs’ Social Abilities.

A range of methods, including training-based, neuro-symbolic, and prompt-based methods, have been proposed to improve the social reasoning of LLMs. Training-based approaches (e.g., SODA (Kim et al., 2023a), SOTOPIA-$\pi$ (Wang et al., 2024)) distill social knowledge from large-scale interactions but struggle with generalization. Neuro-symbolic models (e.g., belief trackers (Sclar et al., 2023) and AutoToM (Zhang et al., 2025b)) offer structure for reasoning but don’t scale well. Prompt-based methods (e.g., SIMTOM (Wilf et al., 2023)) focus narrowly on individual agents, risking the omission of broader social context. Altogether, these methods capture only fragments of social reasoning, revealing the need for structured, general-purpose representations of the full social world state.

(Social) World Models.

Recent advances in LLMs have enabled the development of general-purpose generative world models. However, these world models have primarily focused on representing the physical state of the world (Xiang et al., 2024; Huang et al., 2022; Ding et al., 2024; Liu et al., 2025). Cognitive science research has shown that humans maintain sophisticated models of other agents’ mental states (Sap et al., 2023; Jara-Ettinger & Schachner, 2024). This insight has inspired the theoretical discussion of mental social world (Hinsz, 1995; Ding et al., 2024). As shown in Figure 1, a social world model extends traditional world models to include representations of other agents’ beliefs, intentions, and potential actions. Recent efforts to model social worlds have integrated symbolic representations with neural methods (Dong et al., 2023; Martin, 2021; Zhang et al., 2025b) or developed social simulation systems (Park et al., 2023, 2022). However, these approaches are often constrained to specific domains.

3 Social World Model with S3AP

To overcome the lack of social reasoning in traditional world models (Wong et al., 2023; Ha & Schmidhuber, 2018), we conceptually formalize social world models (§3.1) and introduce a new representation (S3AP) to power these social world models (§3.2).

3.1 Social World Model Formulation

Inspired by $N$-agent Dec-POMDP framework (Bernstein et al., 2013; Nair et al., 2003), we formulate a social world with a state space $\mathcal{S}$, an action space $\mathcal{A}$, an observation space $\mathcal{O}$, a transition function $T:\mathcal{S}\times\mathcal{A}\rightarrow\Delta(\mathcal{S})$, an observation function $\Omega:\mathcal{A}\times\mathcal{S}\rightarrow\Delta(\mathcal{O})$. For $N$ social agents, we define $\mathcal{A}=\mathcal{A}_{1}\times\cdots\times\mathcal{A}_{N}$ and $\mathcal{O}=\mathcal{O}_{1}\times\cdots\times\mathcal{O}_{N}$ as the joint action and observation spaces.

Rather than restricting the agent to conventional world modeling where observations only capture external states, we redefine the observation space to encompass a rich set of social and psychological factors. Specifically, for each agent $i$, the observation space includes both external observations $\mathcal{O}^{\text{ex}}_{i}$ and introspective observations $\mathcal{O}^{\text{in}}_{i}$. The external observations capture information from the environment and other agents, e.g., whether someone exited a room. In contrast, the introspective observations include the agent’s internal mentals states, such as their beliefs, goals, moral values, and emotions. Correspondingly, the agent’s action space expands beyond environmental manipulation to include introspective operations such as recalling memories, reflecting on past actions, and updating beliefs. These expansions enable the agent to act not merely reactively but reflectively, a necessary step for modeling complex social behaviors such as empathy, deception, forgiveness, and norm enforcement (Shen et al., 2024; Su et al., 2025; Forbes et al., 2021).

At time step $t$, each agent $i$ interacts with the social world model by issuing an action $a_{i}^{t}$ and receiving an observation $o_{i}^{t}$. It then makes decision along with its memory $\mathcal{M}_{i}^{t}$ and policy $\pi_{i}:\mathcal{M}_{i}^{t}\times\mathcal{O}_{i}^{t}\rightarrow\Delta(\mathcal{A}_{i}^{t})$. Then a social world model computes:

[TABLE]

Equation (1) predicts other agents’ actions from the social world state, and Equation (2) updates the state given all agents’ actions. Unlike traditional world models, which model passive physical transitions (Ha & Schmidhuber, 2018; Xiang et al., 2023, 2024), this formulation considers other active agents as part of the social world.

3.2 S3AP: Social World State Representation

Building on this formulation, we introduce an LLM-powered general-purpose structured representation of social world states, motivated by the limitations of using raw free-form text alone for social world modeling. Specifically, we propose a protocol to encode such social narratives into a structured representation (i.e., S3AP). As shown in Figure 2, given a free-text narrative describing a social interaction at time $t$, S3AP-parser parses the narrative into a structured representation of a sequence of descriptions for the environment, agents’ observations and actions. We could use either free-form text or special symbols to describe environment state, agents’ observations and actions. For example, <same_as_state> indicates that the agent’s observation is identical to the full environment state. These symbols can be customized and extended to support more complex social interactions for more efficient characterization of $\mathcal{S}^{t}$.

Under the configuration of S3AP, we can encode free-text narratives into a structured representation. This representation serves as an operationalization of the social world state $\mathcal{S}^{t}$ at time step $t$ defined above, up to the information present (or inferable) in the narrative. Given a S3AP representation of a state of the social world $\mathcal{S}^{t}$, a social world model could be a generative model that takes agent’s actions as input and outputs the next environment state, agents’ observations, and other agents’ next actions.

Inducing a social world model in this way offers unique advantages: (1) the protocol provides better structure than pure text narratives with both symbolic and free-form text, enabling more systematic social reasoning (§4). (2) By framing the task as predicting S3AP representation, we can harness the power of LLMs to make the configuration of social world models more efficient and scalable (§5).

4 Representing Social World States for Better Social Reasoning

We first demonstrate the effectiveness of our S3AP representations towards social reasoning by evaluating various LLMs across diverse static third-person perspective social reasoning tasks. We explore the following social reasoning question-answering (QA) tasks:

ToMi & ParaToMi

ToMi (Le et al., 2019) is one of the most important benchmarks for evaluating the theory-of-mind abilities of models. Inspired by the Sally-Anne test, the ToMi dataset evaluates whether models can infer an agent’s belief about an object’s location after a sequence of actions by multiple agents, which may or may not move the object. ParaToMi (Para; Sclar et al., 2023) is a revised version of ToMi (Le et al., 2019) that addresses the limited linguistic diversity of the original by rewording all templates. The resulting dataset is more complex, as actions are expressed in a less straightforward way. For both ToMi and ParaToMi, we randomly sample 600 questions from the dataset. We measure accuracy by whether the model correctly infers the agent’s belief about the object’s location.

HiToM

(Wu et al., 2023) evaluates higher-order theory of mind (ToM) in LLMs, requiring recursive reasoning about others’ beliefs. It extends ToMi by adding agent interactions—such as chatting, deception, and joint attention—beyond simple object movement. The task concludes with a belief inference question about an object’s location. We randomly sample 72 scenarios (100 questions total) and report accuracy based on the model’s ability to infer the correct agent belief.

FANToM

(FAN; Kim et al., 2023b) is a multi-party conversation question-answering dataset designed to test coherent theory-of-mind capabilities. In FANToM, speakers join and leave the conversation while it continues, making participants hold both false and true beliefs. The benchmark includes first-order and second-order theory-of-mind questions about the beliefs of conversation participants. We use 64 sampled conversations from the short version of FANToM, containing a total of 1,086 questions. We report All Qs metric, requiring the model to correctly answer all questions for a given conversation snippet.

MMToM-QA

(MMToM; Jin et al., 2024) is multi-modal question-answering benchmark for theory of mind reasoning, focused on jointly inferring goals and beliefs in everyday object search scenarios. We use the text-only subset (describing search behavior) and evaluate on 300 randomly sampled, balanced questions covering belief (true/false, short/long-term) and goal inference (true/false beliefs, updates, future actions).

4.1 Experimental Setup

We use off-the-shelf LLMs to parse narratives into S3AP representations in the JSON format for ease of computation (parser model). Unlike other ToM methods (Sclar et al., 2023), our representation is general and task-agnostic as we use the same parser prompt across all benchmarks, with minimal adjustments only when benchmarks impose artificial constraints (e.g., ToMi’s assumption that characters automatically know all object locations upon entering a room). Importantly, S3AP parsing is query-independent, creating a general-purpose social world representation for the same narrative without access to downstream questions.

For downstream reasoning tasks, we concatenate the original narrative with its S3AP representation and ask the LLM to answer the question (QA model). Our evaluation spans five state-of-the-art LLMs: GPT-4o, o1, o3-mini, Deepseek-R1 (R1), and Llama 4 Maverick Instruct (Llama 4). Note that we ensure a fair comparison by enforcing that S3AP uses identical source information (i.e., the same narrative) as baselines while providing structured social world representations. We use the Chain-of-Thought (CoT) method as the baseline across all tasks and models (See a more comprehensive baseline comparison in §4.3). To maximize reproducibility, we use temperature 0.0 for all models.

4.2 Main Results Across Tasks and Models

Table 1 shows S3AP consistently outperforms the CoT baseline across all evaluated tasks when averaged across models (e.g., from 0.84 to 0.90 on ParaToMi) 222We use O3 as the parser model here for consistency. Additional experiments with different parsing models show consistent trends (Appendix A.6).. The improvements are especially pronounced in benchmarks requiring complex social reasoning. For instance, FANToM, a benchmark with long, complex multi-agent dialogues, sees the largest average boost of +11.1 points (from 0.39 to 0.50). Surprisingly, even smaller models such as o3-mini and Llama 4 benefit substantially. o3-mini improves on ToMi from 0.86 to 0.96, and Llama 4 shows an increase on FANToM from 0.26 to 0.42. These gains suggest that S3AP helps models disambiguate agent perspectives and maintain coherent mental state tracking even with limited capacity.

To better isolate and understand the effectiveness of the S3AP representations, rather than effects driven by memorization or training contamination, we use ParaToMi for the following ablation studies. ParaToMi requires an access code to obtain, it is less likely to have been included in the training data of most LLMs.

Effect of the parser model

To investigate how the abilities of the parser model affect reasoning, we use different models to generate S3AP representations and apply them to various models for the ParaToMi task. Surprisingly, we find that models generally benefit from the S3AP representations generated by a wide range of LLMs, regardless of the LLMs’ own performance in the social reasoning tasks (Figure 3). For example, despite o3-mini’s 82% lower accuracy on the ParaToMi task, it can generate S3AP representations that boost the o1 model’s accuracy from 84% to 94%. This finding suggests that so-called “social reasoning” may involve two distinct but related components: (1) the ability to track and construct representations of the social world, and (2) the ability to use such representations to answer questions about the mental states of other agents. Importantly, a model’s weak performance on social reasoning tasks does not necessarily imply deficiencies in the social representation construction. This two-part view is consistent with insights from research focusing on the physical world models (Ha & Schmidhuber, 2018; LeCun, 2022; Yerukola et al., 2024).

Influence of the size of the QA model

To understand how QA models with various sizes and reasoning abilities benefit from well-formed S3AP representations (generated by O3), we also conducted experiments on the Qwen3 suite (ParaToMi task; 100 randomly sampled instances), ranging from 0.6B to 80B parameters. S³AP consistently outperforms CoT across all scales. As shown in Table 2, S3AP provides gains even for smaller QA models, while providing larger performance gains for larger models.

Parser Quality and Error Analysis

To understand the quality of the S3AP representations generated by different models, and how it affects the downstream performance, we first manually created S3AP representations for 100 scenarios from ParaToMi serving as the ground truth. We then generate S3AP representations on the same scenarios using various models and evaluated parser quality using an LLM-as-a-judge (GPT-5). Specifically, we asked GPT-5 to rate the quality of the generated S3AP representation by comparing it to the ground truth S3AP representations, which eventually outputs a score on a scale of 0 to 1. While an LLM judge is imperfect, the judge, validated manually, is only used to scale comparisons between a generated parse and the same fixed ground-truth for each instance 333Please refer to Appendix A.3 for more details..

Interestingly, we find that the parser scores correlate with relative improvement in downstream performance (we average the relative improvement over the CoT baseline across all QA models for each parser model in Figure 3): o3-mini (0.707 score, +7.64% improvement), o1 (0.650, +4.44%), GPT-4o (0.620, +6.74%), Llama-4 (0.591, +5.36%), and DeepSeek-R1 (0.490, +3.74%). Notably, smaller models like o3-mini can achieve high parsing scores, suggesting that social parsing capability is distinct from general reasoning ability.

To further understand how the errors in S3AP representations propagate to the downstream performance, we conducted a detailed analysis of 64 randomly sampled misclassified cases from the ParaToMi benchmark (o1 as both the parser and the QA model). We find that the vast majority of errors (79.7%) stem from social context parsing failures (see Figure 4), where models fail to correctly understand and represent the social world state described in the narrative. The remaining errors are split between pure reasoning failures (7.8%) and unspecified cases (12.5%). These findings suggest that building a successful representation of the social world state is the key for successful social reasoning.

4.3 Comprehensive Baseline Comparison

We majorly focus on comparing S3AP with the CoT baseline due to its simplicity and competitive performance on various tasks. To further validate the effectiveness of S3AP, we compare with a wide range of baseline methods in terms of both accuracy and computational efficiency. Specifically, we consider three baseline categories: (1) General prompting: Vanilla LLMs, Chain-of-Thought (CoT), and Few-shot; (2) Specialized ToM methods: AutoToM (Zhang et al., 2025b), which uses automated Bayesian inverse planning for mental state inference. And Thought Tracing (TT) (Kim et al., 2025), which traces mental states by generating and weighting hypotheses based on observations using sequential Monte Carlo-inspired inference. (3) Agentic frameworks: AFlow (Zhang et al., 2025a) and LLM-Debate (Du et al., 2024), which are two recent agentic reasoning frameworks that achieve state-of-the-art performance on various tasks. Those two methods represent the SOTA methods specifically designed for ToM reasoning.

As shown in Figure 5, specialized ToM methods like TT and AutoToM require substantially more LLM calls while achieving lower performance than S3AP. While the number of LLM calls might be acceptable if the generated tokens are minimal (e.g., AutoToM usually generates 1 token per call), these methods become prohibitively expensive for recent reasoning models that generate significant amounts of thinking tokens 444Note that we show the reported AutoToM performance here as we only obtained 48.5% performance on ParaToMi task with AutoToM.. As for the agentic frameworks, S3AP outperforms both AFlow and LLM-Debate while maintaining computational efficiency.

5 Social Interaction with Social World Model

Building on the effectiveness of S3AP as a structured social world representation, We propose a SWM algorithm with S3AP, enabling LLMs to model and predict social dynamics in first-person interactive settings, similar to how humans infer others’ mental states (Frith & Frith, 2006; Forrester, 1971).

Foresee and Act with Social World Model

We propose Foresee and Act, a simple inference-time algorithm that enables agents to simulate the consequences of their actions before committing to them. As shown in Figure 6, the agent first samples a candidate action at each timestep. Then, with the SWM, it simulates how the social state would evolve, including how other agents might interpret the action and how the environment might respond. In our implementation, one LLM predicts the next social world state, and another LLM generates agent’s action and refines it based on the simulated outcome.

5.1 Experiment Setup

We use SOTOPIA (Zhou et al., 2024), the standard benchmark for goal-driven interactive social reasoning. We use the SOTOPIA-hard evaluation set, which features more challenging social scenarios. Each task features two characters with private social goals interacting within a given scenario. During each episode, two agents, a partner and a target of evaluation, play the characters to pursue their hidden goals. In the end, a goal completion score (0-10) is assigned to each agent based on their performance. We evaluate each model on 100 simulations from the SOTOPIA-hard set with GPT-4o as the fixed partner agent and show results for o3, o1, GPT-4.1, and Llama 4 Maverick Instruct (Llama-4) to serve as either the agent, or the social world model, using a one-step SWM rollout to assess the effect of SWM.

5.2 Social World Modeling Results

Figure 7 shows that integrating an S3AP-powered social world model consistently boosts agent performance in social interactions. Interestingly, stronger social agents do not always yield better social world models, for instance, GPT-4.1 outperforms Llama 4 Maverick as an agent (6.01 vs. 4.52), yet their SWM performance is nearly identical (6.34 vs. 6.36). This supports the view in §4.2 that social reasoning involves two parts, and strength in constructing social representations doesn’t imply strength in reasoning over them and acting correctly. Moreover, even when the SWM offers useful information (e.g., o1), the paired agent’s performance can still drop (e.g., with Llama 4 Maverick), highlighting a key challenge: to benefit from social world modeling, agents must be able to effectively incorporate the modeled information into their decision-making. Additional analysis of model pairings and baseline performance can be found in Appendix A.6.1 (Figures 9 and 10).

Analysis on the impact of SWM

To better understand the role of social world models in multi-agent interactions, we evaluate SWM in both cooperative and competitive settings. In the cooperative setting, agents pursue a shared social goal (e.g., Ben and Alice want to identify mutual friends, and they need to chat with each other to find the mutual friends effectively), while in the competitive setting, agents have conflicting goals (e.g., In the barter scenario, Ben and Alice want to trade a TV for a laptop, but they need to negotiate with each other to find a mutually agreeable price). We run 100 simulations per setting using o1 as the agent and o3 as the SWM. As shown in Figure 8, SWM improves performance in both cases, with larger gains in competitive cases. In these cases, the agent can better anticipate and strategically respond to the opponent’s moves. Regradless of the setting, tracking and predicting the intentions and beliefs of other agents is important for successful social interactions, which is the key function of a social world model.

6 Discussion and Conclusion

We define and build social world models through explicit representations of agent mental states, actions, and observations (S3AP). Our approach captures complex social dynamics systematically by automatically transforming free-form narratives into S3AP representations, reducing reporting bias and bridging the gap between raw text and actionable social world models. We achieve SOTA performance on both third-person social reasoning benchmarks and interactive SOTOPIA-hard evaluations by leveraging structured representations and social world models induced from S3AP. We view this work as a first step toward general-purpose social world models for LLMs. Despite these gains, important limitations remain: our approach depends on faithful social parsing and state prediction, and errors in representing implicit norms, private beliefs, or long-horizon dynamics can propagate and mislead downstream reasoning or action. Scaling to deeper interactive settings also raises compute and latency costs, and both parsers and SWMs may inherit biases from underlying LLMs and from culturally contingent social conventions. Furthermore, we majorly evaluate our approach on text-based social reasoning tasks. We believe that our approach can be extended to more complex social reasoning tasks, such as embodied and vision-based social reasoning tasks. Looking ahead, future work should explore improving robustness (e.g., uncertainty-aware parsing and consistency checks), enabling efficient multi-step rollouts for interaction, and training specialized models to generate more accurate SWM efficiently. More comprehensive evaluations that incorporate safety, norm adherence, and human preference will also be essential for deploying social world models in the wild. We envision S3AP as a foundation for building general-purpose social world models that support both social reasoning and interaction across diverse domains.

Acknowledgements

We thank OpenAI for providing credits for running the models in this work. We would also like to express our gratitude to Daniel Fried, Hao Zhu, and Ryan Lin for their valuable feedback on this project. This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR00112490410.

Impact Statement

Our work advances the understanding of social reasoning in AI systems, which has the potential to improve human–AI interaction and enable more socially-aware technologies. At the same time, it is important to consider both societal impacts and reproducibility.

Ethical Considerations, Risks, and Mitigation.

Enhanced social reasoning capabilities could potentially be misused for manipulation or deception. We therefore emphasize transparent development and responsible deployment. Our framework is intended as a research tool to advance scientific understanding rather than for direct deployment in high-stakes applications.

Privacy and Data Considerations.

Our experiments use publicly available datasets and benchmarks. We do not collect new human-subject data, and our experimental protocols follow established ethical guidelines for AI research. We aim to respect privacy and reduce the risk of encoding harmful biases; broader impact considerations are discussed in Appendix A.1.

Reproducibility and Resources.

Complete experimental settings are documented in §4 for third-person reasoning tasks and §5.1 for first-person interactive tasks, including data splits, prompts, and key hyperparameters. Additional implementation details are provided in the appendix. We will release our codebase in the supplemental material, including scripts for data preprocessing and evaluation, along with documentation and usage instructions. Compute resource requirements (hardware specifications and runtime estimates) are reported in Appendix A.7.

Transparency of Claims.

All theoretical results in §3 are accompanied by explicit assumptions and proofs, with detailed derivations in Appendix A.8. Where applicable, we report uncertainty (e.g., error bars) for experiments supporting our main claims (e.g., Figure 7).

Appendix A Appendix

A.1 Limitations

Despite the promising results demonstrated by our S³AP approach, several important limitations must be acknowledged:

Our approach assumes LLM parsers can reliably convert narratives into S3AP structures. In practice, they may struggle with culturally nuanced or ambiguous scenarios, leading to oversimplified or misleading representations (Rao et al., 2025). Experiments were limited to controlled benchmarks (ToMi, ParaToMi, HiToM, FANToM, MMToM-QA, SOTOPIA, and others detailed in the appendix) and focused on accuracy metrics. These settings may not reflect the full complexity of real-world social interactions. Still, consistent improvements across both third- and first-person tasks suggest that S3AP generalizes well. While our Foresee and Act algorithm looks one step ahead, longer rollouts could potentially offer more benefits. However, it also compounds both inference cost and model error: imperfect predicted states may drift over multiple steps, and agents may overfit to artifacts in the simulated trajectory. We left the study for longer rollouts for future work. We don’t explicitly address biases in LLMs, which may carry over into social world models and misrepresent certain groups. Representing mental states also raises privacy concerns. However, the structured format makes such issues more visible and easier to audit than black-box methods. Finally, our method uses predefined templates rather than learning representations from raw experience, which may limit generalization. However, its modular design allows for future integration with learned representation methods as they improve.

A.2 S3AP-Parser Details

Here’s the json schema for the S3AP-Parser.

For all the LLMs powering the parser, we use the 0 temperature if applicable.

A.3 LLM-as-a-Judge Evaluation for S3AP Parser Quality

We evaluate how accurately different parser models generate S3AP representations by comparing a model-produced representation to a manually corrected ground-truth representation on the same ParaToMi scenario. Each evaluation instance contains (i) the original story and question, (ii) a ground-truth S3AP representation, and (iii) a generated S3AP representation from a parser model. This evaluation is designed to compare parsers at scale, but it inherits the limitations of any automated judge. In particular, an LLM judge may overvalue stylistic similarities or under-penalize subtle perspective mistakes. We mitigate this by (a) using a manually corrected reference representation as the anchor for each instance, and (b) focusing the judge on observation fidelity (who saw what, when), which is the primary failure mode in ToMi/ParaToMi.

Step 1: Structural validation.

We first check whether the generated S3AP output conforms to the expected schema (e.g., required fields such as timestep/state/observations/actions are present and well-formed). This yields a structural score in $[0,1]$, where higher is better.

Step 2: Observation accuracy via an LLM judge.

Because the key difficulty in ToMi/ParaToMi is perspective tracking, we focus the judge on whether each agent observes the right things at the right timesteps. Concretely, we provide the judge with the story, the question, the formatted ground-truth S3AP representation, and the formatted generated representation, and ask it to rate observation accuracy (score in $[0,1]$) and provide a brief rationale. The judge is instructed to prioritize whether agents: (i) observe events only when they are present in the relevant location, (ii) do not observe events that occur when they are absent, and (iii) correctly track what they witnessed in prior timesteps. The judge also considers (as secondary criteria) state accuracy and action attribution. We use GPT-5 as the judge and request a JSON response containing a numeric score and reasoning text.

Step 3: Composite score.

We combine the two signals into an overall parser-quality score:

[TABLE]

This weighting emphasizes observation fidelity, which most directly affects downstream theory-of-mind reasoning.

A.4 Experiment Details

For tasks that operate under different assumptions about agent perception and knowledge (e.g., in ToMi tasks, agents are assumed to perceive all events occurring within their physical space), we provide task-specific instructions and one exemplar to guide the encoding process.

A.4.1 Prompt for All Third-person Static Tasks

Here’s the prompt for parsing free-form narratives into S3AP representations.

⬇

Please analyze the following narrative/context.

Context: {context}

Task specific instructions: {task_specific_instructions}

Example analysis: {example_analysis}

Previous attempt had these issues.

Please fix them based on the previous attempt and feedback below:

{feedback}

Follow these format instructions:

{format_instructions}

Here are the task specific instructions for each benchmark:

ToMi

: You are dissecting the TOMI scenarios. The assumptions are that the characters can perceive every scene in their location but not scenes occurring elsewhere. If the agent leaves the location, they cannot perceive the scene in that location anymore. In the agent’s observation, remember to include the objects’ locations if the agents are in the same location as the object.

HiToM

: You are dissecting the HITOM scenarios. You should assume the following: (1) An agent witnesses everything and every movements before exiting a location. (2) An agent A can infer another agent B’s mental state only if A and B have been in the same location, or have private or public interactions. (3) Note that every agent tend to lie. What a character tells others doesn’t affect his actual belief. (4) Agents in private communications know that others won’t hear them, but they know that anyone can hear any public claims. In the agent’s observation, remember to include the objects’ locations if the agents are in the same location as the object.

FANToM

: You are analyzing a social conversation and need to answer a question about it. When the agents leave the conversation, they cannot perceive the conversation anymore untill they join the conversation again. For convenience, you can use <same_as_last_action /> in the state field to indicate that the state is the same as the last action.

MMToM-QA

: You are dissecting the MMToM scenarios. The assumptions are that agents can perceive objects and events only in their current location. When an agent moves to a new location, they can no longer perceive what happens in previous locations. Importantly, agents should not have knowledge about the contents of containers (like fridges, cabinets, etc.) until they directly observe inside them, unless explicitly stated in their prior knowledge. In mental states, clearly represent the agent’s goals, beliefs about object locations, and how these beliefs are updated through observations. In the agent’s observation, include objects’ locations when the agent is in the same location as the objects, but only after the agent has actually observed them.

ConfAIde

: For convenience, you can use <same_as_last_action /> in the state field to indicate that the state is the same as the last action.

Here’s the prompt for question answering:

⬇

Context

{context}

Extra Info

(to help you better understand the meeting)

{extra_info}

Task

{question}

We place S3AP representations in the extra information entry.

A.4.2 Model Configurations

For all experiments, we used the following models:

• •

GPT-4o: gpt-4o-2024-08-06

• •

GPT-4.1: gpt-4.1-2025-04-14

• •

o1: o1-2024-12-17

• •

o1-mini: o1-mini-2024-09-12

• •

o3: o3-2025-04-16

• •

o3-mini: o3-mini-2025-01-31

• •

DeepSeek-R1: together_ai/deepseek-ai/DeepSeek-R1

• •

Llama-4-Maverick: together_ai/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8

• •

Llama-4-Scout: together_ai/meta-llama/Llama-4-Scout-17B-16E-Instruct

For the experiments in Section 4, we used temperature $0.0$ for all the non-reasoning models (reasoning models do not require temperature).

A.4.3 Prompt for Foresee and Act Method

Here’s the prompt for refining the action:

⬇

You are {agent}.

Here is the interaction history between you and the other agent so far:

{history}

Here is your intended action:

{intended_action}

Here is the predicted mental states after you take the intended action

(you should use them to generate better actions for achieving your goal):

{socialized_context_info}

Please generate a refined action

so that you can achieve your (i.e., {agent}’s) goal better.

Please only generate a JSON string

including the action type and the argument.

Your action should follow the given format:

{format_instructions}

A.5 ConfAIde Benchmark Results

ConfAIde

(Mireshghallah et al., 2024) evaluates inference-time privacy in LLMs. We focus on tier 4 meeting summary tasks, where models must include key details while avoiding disclosure of private information to outsiders. This requires reasoning about each person’s knowledge and what information should be shared with different stakeholders. We report the percentage of summaries that satisfy this criterion, using only the meeting summary portion (200 examples) for social reasoning evaluation.

Table 3 shows the performance comparison of models with and without S3AP representations on the ConfAIde benchmark. Our approach demonstrates consistent improvements across most models, with particularly notable gains for models like GPT-4o (from 0.575 to 0.740) and o3-mini (from 0.765 to 0.770). The results indicate that S3AP helps models better reason about information privacy and perspective-taking in meeting contexts.

A.6 Additional Experimental Results

Table 4 shows the performance comparison of models with and without S3AP representations. The S3AP representations is parsed by the o1-based S3AP-Parser.

A.6.1 SOTOPIA Additional Results

Figure 9 presents a detailed heatmap analysis of model performance across different social world model configurations on the SOTOPIA-hard benchmark. The heatmap visualizes how different combinations of agent models and social world models interact, revealing patterns in which pairings yield the strongest performance improvements.

Figure 10 shows the baseline performance of various models without social world modeling support. This serves as a reference point for understanding the improvements gained through S3AP-powered social world models.

A.7 Experiments compute resources

For our experiments, we utilized two main API services for model access:

• •

OpenAI API555https://platform.openai.com/docs/api-reference for accessing GPT-4o, GPT-4.1, o1, o1-mini, o3, and o3-mini models

• •

Together AI API666https://docs.together.ai/ for accessing DeepSeek-R1, Llama-4-Maverick, and Llama-4-Scout models

All experiments were conducted using these cloud-based APIs, eliminating the need for local GPU resources. The API-based approach allowed us to efficiently scale our experiments while maintaining consistent model access across different benchmarks.

A.8 Mapping S3AP to Social World Model Components (Informal)

We provide an informal mapping showing how S3AP aligns with the components of the social world model defined in Section 3.1. This is not a formal completeness proof: any representation is bounded by the information present (or inferable) in the narrative and the parser’s assumptions. The goal here is to clarify what information S3AP makes explicit and how it can be used to operationalize $\mathcal{S}^{t}$ in practice:

State Space $\mathcal{S}$: S3AP approximates the state space through its structured format containing:

• •

Environment state $\mathcal{E}^{t}$ in the state field

• •

Joint observation space $\mathcal{O}^{t}$ in the observations field

• •

Joint action space $\mathcal{A}^{t}$ in the actions field 2. 2.

Observation Space $\mathcal{O}$: S3AP captures both external and introspective observations:

• •

External observations $\mathcal{O}^{\text{ex}}_{i}$ through direct state descriptions and agent actions

• •

Introspective observations $\mathcal{O}^{\text{in}}_{i}$ through the <mental_state> tag 3. 3.

Memory Function $\Psi$: While S3AP doesn’t explicitly store memory, it enables memory reconstruction through:

• •

The sequence of simulation steps that can be used to reconstruct $\mathcal{M}_{i}^{t}$

• •

Special tags like <same_as_last_action> that maintain temporal consistency 4. 4.

Transition Function $T$: The social world model’s transition function is approximated through:

• •

The timestep field that maintains temporal ordering

• •

The sequential nature of simulation steps that captures state transitions

A crucial assumption in our social world model is that agents act independently at each timestep $t$, with no agent having knowledge of others’ simultaneous actions. S3AP supports this assumption through its structured format:

• •

Each agent’s actions are recorded separately in the actions field

• •

The observations field captures only what each agent can observe at the current timestep

• •

The sequential nature of simulation steps ensures that agents cannot access future or simultaneous actions

Furthermore, S3AP satisfies key properties of a social world state, though in an approximate manner:

Completeness: Each simulation step contains all necessary components ($\mathcal{E}^{t}$, $\mathcal{O}^{t}$, $\mathcal{A}^{t}$) to represent a complete social world state at time $t$, though some details may be simplified or omitted. 2. 2.

Consistency: The structured format ensures that observations and actions are consistent with the environment state through:

• •

Special tags that maintain referential integrity

• •

The parser’s ability to infer missing elements through reasoning 3. 3.

Extensibility: The JSON schema allows for additional metadata and future extensions while maintaining the core social world state representation.

Therefore, S3AP provides an approximate but practical operationalization of the social world state as defined in our theoretical framework. While it may not capture every nuance of the theoretical model, it offers a structured and computationally tractable way to represent social interactions, and it makes explicit the key fields required to drive downstream reasoning and simulation.