Empowering Ethical AI: Multi-Agent Systems Design & Real-World Cases

Introduction

Sun Jingwei, a Human‑Computer Interaction researcher at Lenovo and Ph.D. graduate of Peking University’s Institute of Psychology and Cognitive Science, focuses on human‑centered AI agents, AI for good, and practical exploration. He has published nine international papers and an IEEE standard.

1. Development Status and Design Opportunities

The rapid development of large language models has transformed interaction paradigms, moving from single agents to multi‑agent human‑computer interaction. Single agents struggle with complex scenarios, while multiple agents can represent diverse identities and capabilities, coordinating to accomplish complex tasks and delivering deeper, personalized services. Research often uses agents for task execution, but this work emphasizes "AI for good" scenarios, such as assisting elderly users whose cognitive abilities may decline, highlighting agents’ anthropomorphic abilities.

For example, Stanford’s AI Town project created 25 large‑model‑driven agents in a virtual town, each equipped with memory, reflection, and planning modules, allowing them to interact with the environment and each other. Users found the simulated social behavior highly human‑like. A follow‑up study interviewed 1,000 participants, extracted personality traits, and generated agents that could mimic human attitudes and behaviors.

2. Practice Case Sharing

2.1 Multi‑Agent System Supporting Cognitive Diversity

Recommendation algorithms often create "information bubbles" by reinforcing users’ existing beliefs. To break this, we built a system that uses large language models to generate context‑relevant texts and role‑playing agents, encouraging users to explore diverse viewpoints.

Design principles derived from workshops with psychologists, HCI experts, and UX designers:

Provide multiple viewpoints through agents to help users escape information bubbles.

Design mechanisms that stimulate deep thinking about content.

Incorporate gamified elements to increase interaction willingness.

We prototyped a forum‑like platform on the topic of "delayed retirement." AI agents with distinct identities (age, gender, profession, attitude) are inserted into the main feed; users can click to converse with them. The dialogue depth is visualized on a "viewpoint map," rewarding users as they engage with increasingly divergent agents.

2.2 Multi‑Agent System Enhancing Emotional Connection for the Elderly

Traditional reminiscence therapy relies on human companions, which are hard to scale. We created a system with agents representing children, grandchildren, and other roles that converse with seniors around nostalgic content. The system combines multimodal understanding with Retrieval‑Augmented Generation (RAG) to adapt to users’ interests and switch topics when needed.

Lenovo’s naked‑eye 3D displays present old objects in three dimensions, increasing immersion and curiosity. Interaction combines natural language, eye‑gaze, and other multimodal cues.

User trials showed that child‑like agents attracted strong emotional engagement, while senior‑role agents resonated deeply with content, facilitating longer, more meaningful conversations.

3. Design Insights

Agent role settings must be flexible and adapt to individual user profiles; the same agent may be perceived differently by different users.

Interaction logic that feels "human" is more important than visual realism; users value natural, logical, and responsive dialogue.

Agents should be able to converse with each other, as inter‑agent dialogue can feel more natural and entertaining than direct user‑agent interaction.

4. Lenovo AI‑for‑Good Roadmap

Fine‑grained interaction: integrate gestures, facial expressions, and body language to better understand user interest and emotion.

Cognitive‑impairment early screening: use multimodal signals (gestures, expressions, body language) to detect early signs of cognitive decline in seniors.

Device innovation: dual‑screen computers with a 3D display on the upper screen and a touch screen below, offering richer interaction space.

We aim for AI development that serves human well‑being, delivering warmth and care.

Q&A

Q1: Are the prompts and cue phrases for agents manually designed? Can they be automatically expanded?

A1: All dialogue content and prompts are generated by LLMs. In the information‑bubble scenario, agent roles are automatically generated based on the current post/topic, producing five agents with differing attitudes. In the elderly reminiscence scenario, roles are currently preset but can be expanded with generative techniques in the future.

Q2: How do naked‑eye 3D displays perform with elderly users? Did they exceed expectations?

A2: Elderly participants showed strong surprise, often trying to touch the display and asking if they could "feel" it. Even after explanation, the novelty remained. The surprise and immersion sparked greater dialogue desire, confirming our goal of making technology inclusive for all ages.