ToM4AI Workshop Summary - AAAI 2026

Date: January 26, 2026
Location: Singapore EXPO, Garnet Room 213

ToM4AI 2026: Closing Remarks and Key Takeaways

The ToM4AI 2026 workshop concluded with thought-provoking insights from leading researchers exploring the intersection of Theory of Mind (ToM) and artificial intelligence. Here’s a summary of the key themes and reflections shared.

Featured Speaker Highlights

Dr. Maarten Sap emphasized the critical gaps in current large language models around individual, interpersonal, and cultural inference. A central question emerged: how do we build genuine subtext into AI systems? His work on social alignment through public-private knowledge inference (Sotopia project) revealed that humans expertly balance these two types of knowledge to achieve socially strategic goals. He reminded attendees that social life inherently brings social risks that AI systems must navigate.

Prof. Geoff Bird challenged the computer science community with a fundamental question: what can psychology teach us about building better AI? He advocated for MindSpace Theory, which conceptualizes mental representation as existing along continuums rather than discrete categories. Some mental states are inherently harder to represent than others. Critically, he called for abandoning inadequate tests and benchmarks, emphasizing that trait generalization across contexts within individuals forms the core of Theory of Mind.

Prof. Sarit Kraus highlighted the essential role of humans-in-the-loop for providing ethical constraints and complementary advantages. Her research showed that successful agents focus on high-level signals in social communication. Intriguingly, she noted that LLMs appear to develop their own internal subtext, with intra-agent correlation exceeding inter-agent correlation.

Central Questions for the Field

The closing remarks posed several provocative questions for ongoing research:

Do we sacrifice rationality to “feel” more human? Humans operate under pressures that AI systems don’t face: trust, reputation, ostracism, physical threat, death, and companionship needs. These pressures fundamentally shape human cognition and social behavior.

Can we distill ToM benefits without recapitulating human-ness? This raises deeper questions about whether such an objective is even coherent. Potential benefits include energy efficiency, subtextual inference capabilities, and adaptability. However, alignment as a rational objective differs significantly from the messy reality of human existence.

Models vs. engineering goals: Are neural networks intended as models of cognition or purely as engineering solutions? The answer matters because in cognitive modeling we care about bias as revealing truth, while in engineering bias represents a bug to fix. Similarly, we must decide whether substrate-realistic considerations matter or if we only care about instrumental effectiveness with interpretable systems.

Looking Forward

The workshop serves as a reminder that the creativity field’s earlier attempts to computationally replicate human qualities yielded problematic results. As the AI community continues developing Theory of Mind capabilities, balancing scientific understanding with practical engineering goals remains an open challenge.

The ToM-a-thon results and workshop proceedings will be shared via arXiv submission, continuing the conversation around these fundamental questions about building AI systems that can truly understand and navigate social worlds.

Acknowledgments

We are greatful to Edith Cowan University (ECU) for their support to bring students and speakers alike to Singapore and for rewareding our ToM-a-Thon winners.