SOTOPIA-TOM: Theory of Mind and Information Management Evaluation in Multi-Agent Interactions

SOTOPIA-TOM is a multi-dimensional benchmark framework designed to evaluate the ability of LLM agents to manage information in multi-party interaction scenarios with information asymmetry and privacy sensitivity. This framework reveals the persistent limitations of current models in complex coordination scenarios, and Theory of Mind (ToM) interventions have been proven to significantly improve agents' information management performance.

As LLM agents increasingly participate in multi-party interaction scenarios, properly handling information asymmetry (knowing when and to whom to disclose information) has become a key requirement. However, existing benchmarks cannot measure this ability in real multi-party scenarios, which restricts the development of multi-agent systems.

Core Objectives

SOTOPIA-TOM focuses on evaluating agents' navigation capabilities in environments with information asymmetry, privacy-sensitive interactions, and multi-party coordination scenarios (3-5 agents).

Interaction Environment

Supports two communication modes:

• Public communication (broadcast): Share public information and coordinate actions
• Private communication (direct message): Simulate private negotiations

Scenario Design

Includes 160 manually reviewed scenarios covering 8 industry domains, where each agent has unique information fragments and information dissemination depends on different channels.

Evaluation Dimensions

1. Information Sharing Ability: Timely sharing of useful information with appropriate parties
2. Information Acquisition Ability: Proactively seeking missing information
3. Coordination Efficiency: Metrics like task completion time and communication rounds
4. Privacy Protection: Preventing improper disclosure of sensitive information

Composite Metric

The research team integrated the four dimensions into a composite INFOMGMT metric to provide a one-stop evaluation.

Model Coverage

The experiment includes 6 LLM backbone models and three prompting strategies: baseline (standard prompt), privacy enhancement (CoT-privacy), and Theory of Mind intervention (ToM-based).

Key Findings

• Top models like GPT-5 achieved only a 62% INFOMGMT score, with limitations such as information-seeking deficiencies and insufficient privacy decisions.
• ToM intervention had significant effects: GPT-4o's privacy violation rate dropped from 9.9% to 2.2% (↓77.8%), and its INFOMGMT score increased from 15% to 40% (↑166.7%).

Differences in effects of different strategies:

• Standard prompt: Lacks explicit consideration of privacy and coordination
• CoT-privacy: Improves privacy protection but may impair coordination efficiency
• ToM intervention: Achieves the best balance between coordination and privacy

Exposed Limitations

Current LLM agents have persistent deficiencies in complex information-asymmetric coordination, privacy-aware decision-making, and Theory of Mind capabilities.

Platform Value

SOTOPIA-TOM is a scalable testing platform that supports the development of privacy-aware multi-agent systems and research on ToM applications, among others.

Practical Applications

The results can be applied to scenarios such as intelligent customer service, negotiation decision support, privacy-protecting AI, and social robots.

Future Research

1. Enhance Theory of Mind capabilities
2. Optimize privacy-utility trade-offs
3. Multi-agent learning of information management strategies
4. Cross-domain generalization research

Conclusions

Through real scenarios and comprehensive evaluations, SOTOPIA-TOM systematically reveals the information management limitations of LLM agents. The significant effects of ToM interventions indicate that explicit Theory of Mind modeling is a key direction for improving multi-agent system capabilities, providing important standards and directions for related research.