Higher Education Category Submission

420 – GAI Multi-Round & Multi-Agent “Deep Thinking” ——Based on TinyTroupe

Recognizing the limitations of current popular generative AI tools, this project has developed a prototype process of GAI tools, Multi-Round & Multi-Agent "Deep Thinking", based on Microsoft’s open-source collaborative framework, TinyTroupe. The key innovations are as follows:
(1)Multi-agent discussions are conducted at each round of analyzing about the user's problem by deploying multi-agent roles based on the existing LLMs, so as to form a better answer as the basis of the next stage of reasoning until the user is satisfied about the final answer.
(2)Users are motivated to participate in the process of problem analysis since they need to monitor the answers of the multi-agent discussion.
If successfully implemented, this project is expected to help address the diminishment of critical thinking in users caused by over-reliance on generative AI tools.

Upon entering my graduate studies, I noticed an increase in opportunities to apply generative AI (GAI) for interdisciplinary knowledge integration. However, when I posed complex interdisciplinary questions to GAI, its responses were often unsatisfactory—either overly focused on certain aspects or missing some critical points. This year, the "Deep Thinking" feature in GAI tools began displaying its analytical process, which initially seemed helpful for understanding how GAI tools derived answers. Yet, upon further use, I found its utility in refining results remained limited, as it still relied on a single-agent perspective. This led me to hypothesize: if GAI could simulate real-world multi-person discussion, might its generated answers better align with interdisciplinary questions? Therefore, I began to learn about AI agent and its development. That's how I learnt about TinyTroupe and got the idea.
The hypothesis of the project is the multi-round and multi-agent discussion conducted by AI agents can promote GAI performance and better deal with complex educational tasks. The hypothesis is based on following 2 reasons:
(1)Existing multi-agent simulation frameworks, such as TinyTroupe and AutoGen, are available for reference and practical adaptation.
(2)Additionally, educational multi-agent systems have attracted research attention, showing their efficacy in addressing complex educational scenarios and advancing the development of AI in education.

The core functionalities of my prototype solution are:(1)Questioning: Allow users to select the field of their questions while inputting them(by ticking or writing). (2) Deep Thinking: Match AI agents based on the users' query, where the occupation, skills.etc of agents are customized during deployment in advance. At each round of analysis, these agents engage in discussions, with the outcomes serving as the starting point for the next round of analysis. (3)Prompting: Users can interrupt the discussion at any time based on outputs from different agents and propose corrections to the ongoing process. For this project, the performance metrics from users' angle are mainly considered. For example, whether the answers qualify the questions' dimensions/whether the system offers agent roles according to users' need/how would users rate the final answers. In terms of server side, it will relate to more professional types of metrics, like respond time, loading capacity of the server and etc.
The solution is based on current LLMs, like GPT-4O and etc. It means the key point here actually is which type of LLM is used and this determines API that will be used for the prototype's implementation. I plan to validate the market demand by following ways: perform competitive analysis and differentiation positioning aligned with prevailing generative AI tools, evaluate target users' demand criticality through structured surveys and in-depth interviews, and execute a scoped beta-testing to capture actionable feedback.

I participate in Stream 2.

The project offers a technical framework prototype for GAI usage in educational multiple agents application. To demonstrate the innovation and creativity, the project provides the work flow of the GAI Multi-Round & Multi-Agent “Deep Thinking” process. Centered around the core, the project offers a contextual Example to further explain how the Multi-Round & Multi-Agent “Deep Thinking” process works. According to the flow chart, the satisfaction of answers can reflect the effectiveness in addressing user challenges.(Above all can be shown in the uploaded PPT)

I will consider to leverage a architecture with auto-scaling capabilities, dynamically allocating resources based on real-time demand. Bottlenecks such as data throughput limitations and response latency could be reduced by distributed caching (These strategies are considered from server side, thus answers here are just some possible envisaged solutions). Long-term engagement is fostered via adaptive personalization, including integrating user feedback loops and AI agents' feature updates. If this solution is ultimately implemented, it is advisable to consider a plugin-based system to finalize the prototype, which can be convenient for using and updating with evolving needs.

First of all, the over-reliance on GAI is always a concern when mentioning the usage of GAI tools. For my solution, GAI users needs to monitor the answers of the multi-agent discussion, which can motivate users to participate in the process of problem analysis. Further speaking, positively being a role in the process of GAI analysis balance the relationship between GAI and human intelligence, which benefits the development of thinking ability.
Although the solution compensates for existing limitations of GAI in a way, it still necessitates human oversight of its outputs, particularly for ethical considerations. This is actually tied to the inherent ethical risks embedded in LLMs themselves. The key here lies in implementing technical supervision by dedicated personnel, coupled with periodic recalibration of predefined agent roles based on dataset shifts.

Personal Information Collection Statement (PICS):
1. The personal data collected in this form will be used for activity-organizing, record keeping and reporting only. The collected personal data will be purged within 6 years after the event.
2. Please note that it is obligatory to provide the personal data required.
3. Your personal data collected will be kept by the LTTC and will not be transferred to outside parties.
4. You have the right to request access to and correction of information held by us about you. If you wish to access or correct your personal data, please contact our staff at lttc@eduhk.hk.
5. The University’s Privacy Policy Statement can be access at https://www.eduhk.hk/en/privacy-policy.

• I have read and agree to the competition rules and privacy policy.