Metrics, adaptation mechanisms, and shared social memory for Social-AI agency

Create modeling paradigms and metrics that allow Social-AI agents to estimate success in achieving social goals from explicit and implicit signals; develop mechanisms for adapting behavior to achieve single and multiple simultaneous goals; and construct shared social memory between Social-AI agents and other actors to inform algorithms for learning from social signals.

Background

Goal-oriented Social-AI agents must learn from diverse social experiences, often relying on implicit, fleeting, and context-dependent cues rather than frequent explicit feedback. Measuring progress toward social goals in such settings is challenging.

The authors emphasize the need for metrics and adaptation mechanisms that operate over multiple goals and advocate for shared social memory to establish common ground, align social expectations, and support learning from interactions.

References

This challenge leads us to identify several understudied, open questions: How can modeling paradigms and metrics be created for Social-AI agents to estimate how successful they are in achieving social goals, based on explicit and implicit signals? How can mechanisms to adapt behavior towards achieving single goals and multiple, simultaneous goals be developed for Social-AI agents? How can shared social memory be built between Social-AI agents and other actors in interactions, and how can this memory inform algorithms for learning from social signals?

— Advancing Social Intelligence in AI Agents: Technical Challenges and Open Questions (2404.11023 - Mathur et al., 17 Apr 2024) in Section 4, Subsection (C4) Agency and Adaptation, C4 Opportunities and Open Questions