🤖 AI

Social World Models

This paper introduces Social World Models (SWMs) and a novel structured representation formalism called S3AP to explicitly model hidden mental states and social dynamics, significantly enhancing AI agents' ability to reason about and navigate complex social interactions across multiple benchmarks.

Original authors: Xuhui Zhou, Jiarui Liu, Akhila Yerukola, Hyunwoo Kim, Maarten Sap

Published 2026-08-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xuhui Zhou, Jiarui Liu, Akhila Yerukola, Hyunwoo Kim, Maarten Sap

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking through a bustling city. You don't just see buildings and traffic lights; you see people. You notice that the person rushing past you looks worried, the couple arguing on the bench seems angry, and the barista smiling at a customer is being friendly. You are constantly running a secret, invisible movie in your head, guessing what others are thinking, feeling, and planning to do next. This is how humans navigate the world: we build a "mental map" of other people's minds, not just the physical objects around us.

For a long time, computers have been brilliant at mapping the physical world. They know that if you drop a ball, gravity pulls it down, and if you push a door, it opens. But when it comes to people, computers often get lost. They can read a story about an argument, but they struggle to understand the hidden feelings, the unspoken rules, or why someone might lie to protect a friend. They see the words, but they miss the "ghosts" in the machine—the beliefs, intentions, and emotions that drive human behavior. This paper asks a big question: Can we teach AI to build its own mental map of people, just like we do?

The researchers behind this study, published at the COLM 2026 conference, say the answer is yes, but only if we change how we feed information to the AI. They argue that simply dumping a long, messy story into a computer isn't enough. Instead, they propose a new way to organize social information called Social World Models. Think of it like this: if a standard AI reads a novel and tries to guess the ending, it's like trying to navigate a dark room by feeling the walls. A Social World Model, however, turns on the lights. It breaks the story down into a structured checklist: Who is here? What do they see? What are they secretly thinking? What did they just do?

To make this work, the team invented a special format called S3AP (Structured Social Simulation Analysis Protocol). Imagine S3AP as a translator that takes a chaotic, free-flowing story and turns it into a clean, organized dashboard. Instead of a paragraph saying "Isla was nervous because her mom saw her," S3AP breaks it down into clear fields: Agent: Isla; Emotion: Nervous; Observation: Mom saw me; Action: Hiding the game controller. This structured data acts as a "social blueprint" that the AI can actually reason with.

The paper tests this idea in two main ways. First, they gave the AI a bunch of social puzzles (like figuring out who knows what in a group chat) and asked it to solve them. The results were impressive: when the AI used the S3AP blueprint, it got significantly better at solving these puzzles. On one tricky test called FANToM, the AI's performance jumped by 51% compared to its previous best attempts. Even more interestingly, they found that a smaller, less powerful AI could act as the "translator" to create the blueprint, and then a smarter AI could use that blueprint to solve the puzzle even better than if it had tried to do it alone. This suggests that the skill of organizing social information is different from the skill of solving the problem, and both are needed.

Second, the researchers let the AI play a game where it had to interact with another AI in real-time, like negotiating a price or making a new friend. They gave the AI a "forecaster" tool: before the AI said or did anything, it would use its Social World Model to simulate, "If I say this, how will the other person feel? What will they do next?" This simple step of pausing to think ahead made the AI much more strategic. In competitive situations, like bargaining, the AI improved its success rate by up to 18%.

The paper suggests that by giving AI a structured way to track not just what people do, but what they think and feel, we can help them navigate social situations with much more intelligence. It's not a magic fix that solves every social problem instantly, and the researchers admit there are still challenges, like making sure the AI doesn't get the feelings wrong. But the findings strongly suggest that when AI stops just reading words and starts building a mental model of the people behind them, it becomes a much better social partner.

Technical Summary: Social World Models

Problem Statement

Current AI systems struggle to navigate complex social interactions because they lack explicit computational frameworks for reasoning about unobserved social dynamics. While humans intuitively simulate unspoken intentions, beliefs, and evolving social states, AI models typically rely on static, free-text narratives. These narratives suffer from inherent limitations:

Reporting Biases: They often omit explicit mentions of mental states, perspectives, and partial subjective experiences, focusing only on salient events.
Lossy Representations: Traditional inputs fail to capture the recursive nature of social reasoning (e.g., "I think that you think that I...").
Lack of Structure: Existing world models primarily track physical states (location, temperature) but ignore the "social fabric" of interactions, such as agents' beliefs, desires, and intentions.

Consequently, AI systems lack the mechanisms to effectively model others' perspectives or navigate social contexts where information is implicit and incomplete.

Methodology

The paper introduces Social World Models (SWMs), a conceptual framework that decomposes social intelligence into two dissociable components: constructing structured representations of social world states and reasoning over them. To operationalize this, the authors propose S3AP (Structured Social Simulation Analysis Protocol).

1. S3AP Representation

S3AP is a minimal, sufficient, and inspectable structured representation designed to bridge lossy free-form narratives and the rich data required for social reasoning. It encodes:

State: The overall environment state.
Observations: What each agent perceives (including external observations and introspective mental states like beliefs, goals, and emotions).
Actions: The actions taken by each agent.
Mental States: Explicit tracking of hidden variables (intentions, beliefs) that are often absent in raw text.

The representation is generated via an S3AP-Parser, an LLM-powered tool that converts free-text narratives into structured JSON-like formats. This allows the system to explicitly state social information before the reasoning model processes it.

2. Social World Model Formulation

The authors formalize the social world using an N-agent Dec-POMDP (Decentralized Partially Observable Markov Decision Process) framework. Unlike traditional world models that model passive physical transitions, the SWM formulation:

Redefines the observation space to include introspective observations (internal mental states).
Expands the action space to include introspective operations (e.g., recalling memories, updating beliefs).
Treats other agents as active components of the world state, predicting their actions ( $p(A_{-i} | S_t)$ ) and updating the global state based on joint actions.

3. Inference-Time Algorithm: Foresee and Act

For interactive settings, the paper introduces Foresee and Act, a one-step lookahead algorithm. Before an agent commits to an action, the SWM:

Samples a candidate action.
Predicts the next social state (including how other agents interpret the action and their resulting mental states).
Refines the agent's action based on this simulated outcome.

Key Contributions

Formulation and Representation: The paper defines SWMs as computational models tracking environment, observations, actions, and evolving mental states. It operationalizes this via S3AP, a structured protocol that makes implicit social dynamics explicit.
General Parser and Systematic Study: The authors developed an S3AP-Parser and evaluated it across five third-person social reasoning benchmarks (ToMi, ParaToMi, HiToM, FANToM, MMToM-QA).
Interactive Social World Model: The paper introduces the "Foresee and Act" algorithm, enabling agents to predict social consequences and refine actions in a first-person interactive setting (SOTOPIA).

Experimental Results

Third-Person Reasoning Benchmarks

The authors evaluated S3AP + Chain-of-Thought (CoT) against standard CoT across five models (Llama 4, GPT-4o, o3-mini, DeepSeek-R1, o1) and five tasks.

Performance Gains: S3AP + CoT outperformed standard CoT in 30 out of 35 model-task combinations.
Significant Improvements:
- FANToM: Achieved a ~51% relative improvement over OpenAI's o1 (from 0.415 to 0.623 accuracy, a +20.8 percentage point absolute gain).
- ParaToMi: Improved o1 accuracy from 83.5% to 93.2% (a +9.8 percentage point gain).
- Cross-Model Transfer: A less capable parser (o3-mini) generated representations that improved a stronger reasoning model (o1) more than o1's own representations did. This suggests that representation construction and reasoning are distinct capabilities.
Ablation Findings:
- The gains are driven by the explicit modeling of hidden mental states, not merely the JSON serialization format (bullet points performed similarly to JSON).
- Errors in the final output were primarily caused by social context parsing failures (79.7% of errors), rather than pure reasoning failures, highlighting the importance of accurate representation construction.

Interactive Social Interaction (SOTOPIA)

In the SOTOPIA-hard benchmark (goal-driven social interaction):

Overall Improvement: Agents equipped with an S3AP-powered SWM showed up to ~18% relative improvement (a +1.10 point gain on a 0-10 scale) in goal completion scores for the best pairing.
Cooperative vs. Competitive: The SWM yielded gains in both settings, with notably larger improvements in competitive scenarios. This suggests the model effectively utilizes information about the interlocutor's goals and information asymmetry.
Decoupling: Direct agent performance and the ability to utilize a social-state prediction did not always align; a weaker agent could benefit significantly from a strong SWM.

Significance and Claims

The paper argues that social intelligence requires decomposing into representation construction and reasoning. The primary significance of this work lies in demonstrating that:

Explicit Representation Matters: Structuring implicit social dynamics (beliefs, intentions) into a formal representation (S3AP) significantly enhances LLM performance, even surpassing the capabilities of state-of-the-art reasoning models (like o1) when used as a reasoning aid.
Separation of Concerns: The ability to construct a social representation and the ability to reason with it are dissociable. A model that is not the best at answering a task directly can still produce representations that enable a stronger model to reason more effectively.
Foundation for Social AI: Just as physical world models help agents anticipate physical consequences, Social World Models enable AI systems to reason about and navigate social interactions by simulating unspoken dynamics and predicting interlocutor behavior.

The authors position S3AP as a foundational step toward more socially aware AI, noting that while current limitations exist (parsing dependence, text-only scope), the framework provides a structured path for future research in robust, multi-turn social interaction.