What role does Theory-of-Mind capacity play in LLM multi-agent collaboration?

Ascertain the role that theory-of-mind reasoning capacity of large language models plays in collaboration among agents within multi-agent systems.

Background

The authors find that prompting LLM agents to consider others’ actions can increase goal-aligned coordination in their experimental task, while lower-capacity models (e.g., smaller LLMs) show limited benefit from such instructions. This suggests ToM-like capabilities might be important, but the extent and mechanisms remain unsettled.

A general understanding of how ToM capacity influences emergent coordination, role differentiation, and overall collaboration quality across varied tasks and model families is an open area of inquiry.

References

While these principles likely apply to LLM-based collectives given the universality of the integration-segregation tradeoff, it remains unclear whether agents develop differentiated roles, whether such roles complement each other, how to steer it with prompts, and what role ToM capacity of models plays for collaboration.

— Emergent Coordination in Multi-Agent Language Models  (2510.05174 - Riedl, 5 Oct 2025) in Related Work, Section: Related Work

How to incorporate theory of mind into every decision the assistant makes remains an open question for future work.

— Collaborative Theory of Mind: Learning to Self-Revise by Anticipating User Responses  (2610.12872 - Zeng et al., 9 Oct 2026) in Section 6, Conclusion and Discussion