Plural Voices, Single Agent: Towards Inclusive AI in Multi-User Domestic Spaces

TL;DR

This work introduces the Plural Voices Model (PVM), a single-agent domestic AI framework that dynamically negotiates the needs of diverse household members through real-time value alignment. Implemented as the AgoraNest app with the Agora-4B model, PVM emphasizes privacy-preserving memory, adaptive safety scaffolds, and participatory co-design, and it demonstrates improved compliance, fairness, and safety over multi-agent baselines in preliminary evaluations. The study provides design guidelines (participatory design, transparency, adaptive modalities, IDEAS safeguards, human-AI partnership) and an extensive methodology combining diverse datasets, synthetic curriculum training, and multi-axes LLM evaluation. Collectively, the findings offer a path toward inclusive, user-centered agentic AI in plural domestic spaces, while acknowledging limitations of synthetic data, short-term studies, and potential biases in evaluation pipelines. The work also contributes open-source resources (Agora-4B, AgoraNest) to support reproducibility and further research in inclusive household AI.

Abstract

Domestic AI agents faces ethical, autonomy, and inclusion challenges, particularly for overlooked groups like children, elderly, and Neurodivergent users. We present the Plural Voices Model (PVM), a novel single-agent framework that dynamically negotiates multi-user needs through real-time value alignment, leveraging diverse public datasets on mental health, eldercare, education, and moral reasoning. Using human+synthetic curriculum design with fairness-aware scenarios and ethical enhancements, PVM identifies core values, conflicts, and accessibility requirements to inform inclusive principles. Our privacy-focused prototype features adaptive safety scaffolds, tailored interactions (e.g., step-by-step guidance for Neurodivergent users, simple wording for children), and equitable conflict resolution. In preliminary evaluations, PVM outperforms multi-agent baselines in compliance (76% vs. 70%), fairness (90% vs. 85%), safety-violation rate (0% vs. 7%), and latency. Design innovations, including video guidance, autonomy sliders, family hubs, and adaptive safety dashboards, demonstrate new directions for ethical and inclusive domestic AI, for building user-centered agentic systems in plural domestic contexts. Our Codes and Model are been open sourced, available for reproduction: https://github.com/zade90/Agora

Figures (10)

• Figure 1: We designed personal AI agentic voice,app, and video assistant to support children, older, Neurodivergent adults’ self-management in 5-stages: stage 1— recruitment of interested participants and their consent; stage 2—interviews with all the user types and getting familiar with their individual needs; stage 3— Model training based on necessary open-source data and initial prototype development; stage 4— co-design workshops and conflict scenarios discussion and refining the prototype; and stage 5— Human based and LLM based validation based on the user satisfaction for all the participants. All activities with participants were consented and compensated for their participation.
• Figure 2: Multi-agent architectures in home automation face challenges with concurrent scenarios, leading to model confusion, hallucinations, and potential emotional harm to the user.
• Figure 3: The Agora-nest is a single-agent architecture designed to quickly & effectively resolve both concurrent and individual queries in home automation with its memory and ethical reasoning capabilities.
• Figure 4: Interview findings and their underlying causes are presented, followed by a summary of the design requirements. We demonstrate how these requirements were implemented in our initial prototype based on user's requirements.
• Figure 5: Combined responses of all the models we tested against our model Agora-4B and here is a query and response chosen in case of child user for single agent system. As we can see Red in the question represents actual need of the user, and in the answer Red represents the actual desired reply, Blue represents safety and feedback based further actions by the agent, Green+italics represents vague answers and hallucinations