Computational Theory of Mind for Effective Hybrid Intelligence

The rapid growth of Artificial Intelligence (AI) has led to increasing interaction between humans and autonomous agents in complex, dynamic environments. While AI systems excel at computational and data-driven tasks, they often lack the social understanding and adaptability required for effective collaboration with humans. Hybrid Intelligence offers a paradigm where humans and machines combine their complementary strengths to tackle challenges together. A central element for such collaboration is the computational implementation of Theory of Mind (ToM) reasoning: the ability to reason about the mental content of others such as their beliefs, desires, preferences, and goals. This thesis investigates how computational agents can leverage abstraction-based ToM to enable effective and human-aligned collaboration. In this work, abstractions are framed as human-inspired decision-making heuristics such as trust. By reasoning with these abstractions, agents can model complex social dynamics of collaboration. As a result, the agents can reduce reasoning complexity and enhance interpretability, enabling them to make well-aligned and adaptive decisions while interacting with humans in dynamic, multiagent settings. The thesis is structured around four research questions, each addressed in a dedicated chapter. Chapter 1 introduces the ToMA framework, which formalizes ToM using abstractions as high-level constructs for reasoning about others’ mental attitudes. Through a healthcare teamwork scenario, it shows how abstraction-based reasoning allows agents to interact with humans effectively without tracking every belief and knowledge state individually. Chapter 2 embeds value-sensitive reasoning into ToM and focuses on privacy as a key social value. It presents privacy-aware agents operating in multiagent settings and demonstrates through simulations that these agents better align group decisions and mitigate privacy conflicts compared to baseline strategies. Chapter 3 proposes modular architecture that maintains three forms of internal consistency: abstraction consistency, ToM consistency, and goal consistency. This architecture enables scalable and adaptive ToM reasoning, validated through simulations. Chapter 4 introduces relational operators for managing interactions among multiple abstractions, such as trust, respect, and suspicion. A privacy-sensitive case study illustrates how these relational structures can support human-agent collaboration by fostering trust. Collectively, these contributions advance the field of Hybrid Intelligence by bridging cognitive theories of social reasoning with computational models. The ToMA framework establishes a formal and interpretable foundation for abstraction-based ToM. The development of privacy-sensitive agents demonstrates how value-aligned reasoning can address real-world challenges in multiagent systems. The modular architecture provides mechanisms for maintaining consistency and adaptability over time, while the relational reasoning framework captures the complex and interconnected nature of social dynamics. This work lays the groundwork for developing AI collaborators that can effectively partner with humans in Hybrid Intelligence settings, with potential applications across domains such as healthcare and privacy management systems.