Introduction
Personalization in AI has traditionally focused on the individual. However, as AI agents move into shared spaces like smart homes and collaborative workplaces, they must navigate the preferences of multiple users simultaneously. This shift introduces the critical challenge of conflict resolution. For instance, in a smart home, one resident might prefer a cooler temperature for productivity while another prefers warmth for comfort. Standard "monolithic" Large Language Models (LLMs) often struggle in these multi-user environments, frequently failing to retrieve all relevant constraints or hallucinating resolutions that do not satisfy the group's collective needs.
The MAP Framework
To address these challenges, the researchers developed MAP (Multi-Agent for Multi-user Personalization). MAP moves away from a single-agent approach, instead employing a team of specialized LLM agents organized around a workflow inspired by conflict resolution theory.
The framework operates in three distinct stages:
- Reflection: The system aggregates preferences, history, and schedules from all users.
- Analysis: The system synthesizes this data to create a plan, explicitly identifying conflicts and proposing resolutions based on pre-defined priority rules.
- Feedback: Users review the plan and provide iterative adjustments.
The architecture consists of three core agents:
- The Planner: Acts as the high-level interface, managing user communication and final reasoning.
- The Rule Manager: Decomposes complex queries into granular, executable sub-queries.
- The Rule Retriever: Utilizes Retrieval-Augmented Generation (RAG) to fetch relevant user data from a vector database.
By delegating retrieval to specialized agents, the Planner remains focused on high-level reasoning, preventing context-window overload and improving overall accuracy.
Evaluation Scenarios
MAP was evaluated across three "Everyday Scenarios" simulating common multi-user friction points:
- Workplace Scheduling: Balancing room preferences and prioritizing client meetings.
- Assistive Care Robot: Managing wake-up calls and deliveries for care facility residents.
- Smart-home Temperature Control: Regulating communal areas based on housemate schedules and cost-saving goals.
Each scenario involved 60 rules and 12 specific conflicts, with performance measured against a "ground truth" reference solution.
Key Results
The multi-agent approach significantly outperformed monolithic LLM baselines in both information retrieval and conflict resolution.
- Retrieval Accuracy: MAP achieved approximately 80% accuracy in retrieving relevant user preferences, whereas monolithic models hovered around 50%.
- Conflict Resolution: MAP successfully identified and resolved conflicts at a rate of 60-70%, compared to as low as 30% for the monolithic approach in complex scenarios.
- User Sentiment: A study with 12 participants revealed that users found the system highly effective at reducing the "cognitive load" of group mediation. Participants particularly valued the "Analysis" stage, which provided transparency into the AI's decision-making process.
Discussion and Future Work
The study highlights that modularity beats scale; splitting tasks into specialized roles is more effective than relying on a single large model. Furthermore, users expressed a surprising willingness to delegate the "mentally draining" task of conflict mediation to an AI, provided the system remains transparent.
However, limitations remain. Retrieval failures still occur, and the system's scalability to larger groups (e.g., large offices) has yet to be tested. Future work may involve incorporating Formal Verification to ensure strict adherence to hard constraints and adding "Critic" agents to further validate the AI's proposed resolutions.
Summary updated