Human–AI InteractionRecursive Self-ImprovementInstitutional Design
AI can already discover new mathematics on its own, yet stalls on questions of policy and society. I study that gap — and how fast, hard-to-fake validators for slow-verification domains can close it, keeping AI progress both genuine and trustworthy.
My work spans computational social science, AI self-improvement, and governance design. One thread runs through all of it: in domains where outcomes are slow and hard to verify, how do we build the fast validators — and the institutions — that keep AI progress both genuine and trustworthy? The three directions below move from describing human–AI interaction, to the methods that let AI make real progress in such domains, to designing the institutions and safeguards around it.
Empirical and computational study of how people interact with AI systems, and how those interactions aggregate into social-scale dynamics. Spans mixed-methods behavioural experiments (N = 86) on autonomy, trust, and delegation; NLP analysis of cross-cultural AI sentiment; and dynamical-systems modelling of how AI controversies propagate through opinion networks.
Recursive self-improvement delivers real gains only where fast, hard-to-fake validators exist, as in mathematics and code. I study how to sharpen slow-validation domains such as policy by building fast surrogate validators — agent-based and dynamical-systems models, structural invariants — so that AI can propose bold hypotheses and have them tested safely.
Turning sharpened validators into governance: designing accountable institutions for AI-enabled public resource allocation, grounded in mechanism design, Bayesian Truth Serum, and agent-based stress-testing. Targets two pathologies of autonomous delegation — moral responsibility diffusion and delegation blindspots.
On the safety side: benchmarking LLM-based safety guards across Southeast Asian languages, and fine-tuning lightweight multilingual safety classifiers as fast-response decision models.
Peer-reviewed work, then working papers (under review or on arXiv), then unpublished writing samples. Full CV ↗
Comparison of hand tracking-based and controller-based interaction in a consumer virtual reality game
A controlled VR study comparing two interaction modalities — hand tracking versus traditional controllers — in a consumer game setting. Contributes empirical evidence on how input modality shapes user behaviour, presence, and performance.
Virtual RealityHuman-Computer InteractionInteraction Modalities
KITA AI: A Multi-Agent LLM System for Pluralistic Policy Deliberation
Multiple LLM agents, each grounded in a distinct demographic stakeholder persona and conceptual framework, deliberate on policy scenarios. Rather than converging on a single recommendation, KITA AI surfaces who is affected and why, and treats non-convergence as a first-class, explainable output. Jointly developed by UNU Macau, AI Singapore, and the University of Pretoria.
Multi-Agent LLMsPolicy DeliberationPluralismSDGs
Stability of AI Governance Systems: A Coupled Dynamics Model of Public Trust and Social Disruptions
Derives closed-form spectral stability conditions (ρ(J2n) < 1) for a bidirectionally coupled Hawkes–Friedkin–Johnsen trust dynamics system. Demonstrates irreversible governance collapse under echo-chamber topologies and identifies tipping-point thresholds for institutional intervention.
AI GovernanceDynamical SystemsPublic TrustHawkes Process
Unpublished manuscripts and coursework, shared as samples of my writing.
Can freedom lead to trust & satisfy? How user autonomy influences trust perception in LLM-powered conversational agents
A mixed-methods study (N = 86) on how perceived user autonomy relates to trust and satisfaction in LLM-powered conversational agents.
Read PDF ↗The Disinformation Campaigns about the Israel-Palestine Conflict in the Digital Era
Analyses the systemic impact of generative AI on state-sponsored computational propaganda through the lens of Strategic Narrative Warfare theory.
Read PDF ↗Research, policy, and industry roles across Singapore, the UK, and China. Full CV ↗
under Dr. William Tjhi
Supervised research within the PhD-level Mathematical Economics course, taken as an MSc student
under Prof. Yvonne Rogers (UCLIC Director)
Internship · venture investing
Research systems, open-source code, and tools.
A modular multi-agent LLM system in which stakeholder agents debate a policy proposal, raise objections, and concede points, while independent report engines assess it against SDG indicators and regional and philosophical frameworks. Disagreement is kept as an explainable output rather than flattened into consensus, and humans can pause and correct the debate. As co-first author, I built its backtesting, human-review checkpoint, and self-expanding simulation module.
Multi-Agent LLMsHuman-in-the-LoopAgent-Based SimulationSDGs
A daily human–AI brainstorm loop. Each morning it proposes four research ideas (high-fit, adjacent-stretch, contrarian, and random mutation), learns your taste from how you re-rank and comment on them, and consolidates durable lessons into a versioned “taste skill.” Model-agnostic, with built-in novelty and anti-repetition guards and an optional self-evolution step gated by held-out backtests. The asset is the append-only feedback log that compounds over months.
Adaptive AgentsHuman-in-the-LoopLLM ToolingPython
I hold an MSc in Information Studies from Nanyang Technological University and a BSc in Psychology & Language Sciences from University College London. My research sits at the intersection of computational social science, mechanism design, and AI governance.
My work combines formal modelling with behavioural experiments: I build coupled dynamical systems to characterise how AI controversies erode public trust, and I run empirical studies to understand how people delegate to — and hold accountable — AI-powered agents. I am also developing governance frameworks that address moral responsibility diffusion in AI-enabled public resource allocation.
I am currently a Research Associate at AI Singapore (under Dr. William Tjhi), where I co-develop KITA AI, a multi-agent policy deliberation system built with UNU Macau, and evaluate multilingual AI safety guards for Southeast Asian languages. I have previously conducted research at the UCL Interaction Centre (under Prof. Yvonne Rogers), NTU School of Economics, and the China Development Research Foundation, with industry experience in machine learning deployment, quantitative modelling, and deep-tech venture investing.
Aug. 2024 – Jul. 2025
MSc Information Studies
Nanyang Technological University
Singapore
Sep. 2021 – Jun. 2024
BSc Psychology & Language Sciences
University College London
London, United Kingdom
Formal & Quantitative Methods
AI Evaluation & Governance
AI Safety Evaluation
Empirical & Mixed Methods
Machine Learning & Programming
Technology & Investment Analysis
Happy to discuss research ideas, potential collaborations, or questions about my work. Based in Singapore.