Fair Agents: Balancing Multistakeholder Alignment in Multi-Agent Personalization Systems

TL;DR

This paper proposes a conceptual framework for designing fair multi-agent personalization systems that balance stakeholder objectives using alignment, aggregation, and evaluation methods, demonstrated through a tourism case study.

Contribution

It introduces a novel framework integrating stakeholder alignment, social choice-based aggregation, and stakeholder-centric evaluation for multistakeholder multi-agent personalization systems.

Findings

Framework effectively balances stakeholder goals in a tourism scenario.

Uses social choice theory for fair decision aggregation.

Discusses datasets and fairness tensions across domains.

Abstract

LLM agents are increasingly used for personalization due to their ability to communicate directly with users in natural language, integrate external knowledge bases, and negotiate with other (possibly human) agents. Especially in multistakeholder AI systems with multiple distinct objectives, LLM agents are used to independently optimize for each stakeholder's goals. Here, stakeholder alignment is essential to identify and map these goals to provide LLM agents with quantifiable objectives. Plus, the way in which the outputs of the LLM agents are aggregated is fundamental to ensuring fair outcomes for all agents and, therefore, stakeholders. In this work, we identify open research challenges and propose a conceptual framework for designing fair multi-agent multistakeholder personalization systems that balance competing stakeholder objectives. Our framework integrates (i) methods to align stakeholder objectives and LLM agents, (ii) aggregation strategies, e.g., based on social choice theory, to form fair collective decisions, and (iii) stakeholder-centric evaluation procedures for both individual and collective agent behavior. We showcase our framework through a tourism use case and discuss possible applications in other domains, such as education and healthcare. Finally, we discuss domain-specific fairness tensions and review datasets for evaluating multistakeholder fairness and multi-agent personalization systems.