Multi-perspective reasoning for Social-AI agents

Develop models that enable Social-AI agents to perceive and reason over concurrent, interdependent perspectives of multiple actors during interactions; determine whether a single joint model or multiple actor-specific models more effectively represent social phenomena across an interaction; and design mechanisms to efficiently and accurately update agents’ perceptions of other actors’ perspectives during intermittent interactions over time.

Background

In social interactions, each actor maintains a subjective, evolving perspective influenced by roles, settings, norms, and attributes. These perspectives are interdependent, with each actor’s view affecting and being affected by others over time.

The authors highlight the need for modeling paradigms that capture this multi-perspective dynamics, spanning theory-of-mind and counterfactual social reasoning, and call for benchmarks in naturalistic settings to assess such capabilities.

References

This complexity leads us to identify the following open questions: How can researchers create models for Social-AI agents to perceive concurrent, interdependent perspectives of actors during interactions? To what extent would a single, joint model be more effective than multiple models (e.g., one for each actor) to represent social phenomena across an interaction? When interactions occur intermittently over time, how can models efficiently and accurately adjust agents' perceptions of other actors' perspectives?

— Advancing Social Intelligence in AI Agents: Technical Challenges and Open Questions  (2404.11023 - Mathur et al., 2024) in Section 4, Subsection (C3) Multiple Perspectives, C3 Opportunities and Open Questions

Our training environments are primarily dyadic; the zero-shot FANToM transfer in Appendix~\ref{app:fantom-transfer} provides initial multi-party evidence, while end-to-end multi-party training with participant-specific latent updates remains future work.

— Mental-Models for Multi-Agent Systems  (2610.12453 - Gani et al., 8 Oct 2026) in Section 5, Conclusion, paragraph 'Limitations and future work'

Datasets designed around such interaction would address a dimension that current embodied datasets largely leave open.

— Towards a Data-Centric Understanding of Spatial Reasoning in Multimodal Large Language Models  (2610.19108 - Mallahnezhad et al., 13 Oct 2026) in Section 6, Future Directions, paragraph “Human-centred settings”