Wilson is a scholarship recipient of the ARC Centre for Automated Decision-Making and Society.

Each year the Centre offers a select number of PhD scholarships across our nodes to engage and prepare the next generation of researchers to make world-leading contributions in an increasingly engaged and transdisciplinary research environment.

Thesis Title
Advancing Multi-Stakeholder Decision-Making through LLM-Driven Generative Agent Simulations

Research Description
In an era of unprecedented complexity and rapid change, traditional decision-making processes often fall short in addressing multifaceted societal challenges. The GenAISim research project aims to develop a cutting-edge hybrid decision-making system that leverages the power of Large Language Models (LLMs) and multi-stakeholder interactions through generative agent simulations.

At the core of GenAISim is the development of a novel suite of generative and data-driven simulations designed to depict current and future urban scenarios across various domains, including mobility, urban policymaking, and health. By establishing a new “simulation in the loop” paradigm within a multidisciplinary socio-technical framework, this research seeks to revolutionise automated decision-making tools for complex, context-rich challenges.

The project’s primary focus areas include:

• LLMs on HCI/Value Alignment: Investigating the intersection of LLMs and Human-Computer Interaction to ensure that AI-driven decision-making aligns with human values and preferences.
• LLM-based Dialogues: Developing sophisticated dialogue systems that can accurately simulate multi-stakeholder interactions and communication patterns for decision-making objectives.
• Multi-agent Personification: Creating realistic and diverse agent personas that reflect the complexity of real-world stakeholders, including their objectives, priorities, constraints, and incentives.
• A significant portion of this research will concentrate on synthetic data generation for LLMs, addressing the current lack of comprehensive personification data. This includes creating dialogues, developing persona system prompts, and modelling human preferences and routines using LLMs. Additionally, the project will involve designing and implementing evaluation benchmarks to assess the quality and fidelity of LLMs as human simulators.

By exploring LLM agent-based synthetic data generation techniques, developing robust frameworks for hypothesis testing, and investigating diverse stakeholder settings, this project aims to push the boundaries of AI-driven decision-making. The project will also dive into the ethical implications and practical considerations of using LLM agents as substitutes for human decision-makers in various contexts.

Wilson Wongso is a machine learning engineer specializing in Natural Language Processing (NLP), Automatic Speech Recognition (ASR), and Text-to-Speech (TTS). He started his PhD at CRUISE-UNSW under Professor Flora’s supervision in August 2024. He holds a Bachelor of Computer Science degree from Binus University and has completed study abroad coursework at Monash University in Melbourne. With professional experience as a Machine Learning Engineer at Bookbot since 2021, Wilson has worked on various cutting-edge technologies, including real-time speech recognition and text-to-speech models for multiple languages. He is also an active contributor to the open-source NLP community, with several projects on HuggingFace. His research interests include simulating human decision-making with language models and advancing NLP for low-resource languages.

Supervisors
Prof Flora Salim, UNSW