Reading the gap between what firms say and what they do.

I am an interdisciplinary researcher working across accounting, computer science, and artificial intelligence. My research centres on whether judgements about corporate disclosure can be checked. When a system reports that a company has disclosed something, that claim should be verifiable against the reporting standard it was measured by. Working between accounting and computer science, I build evidence-grounded methods for ESG disclosure assessment that represent what each standard actually requires, retrieve the passages that could satisfy it, and test whether the evidence meets those conditions rather than merely mentioning the right words.

I am currently pursuing a Master of Commerce in Business Analytics at the University of New South Wales. I previously earned a Bachelor of Business Administration (Honours) in Accounting from Hong Kong Baptist University, having studied at Beijing Normal-Hong Kong Baptist University in Zhuhai. My undergraduate thesis, supervised by Dr. Man Hung Alvin Cheng, examined the impact of ESG disclosure on corporate investment efficiency. That accounting and ESG background now underpins computer-science research on knowledge representation, retrieval, and the evaluation of language-model systems.

My research interests are LLM evaluation, retrieval-augmented generation, knowledge graphs, and ESG and sustainability disclosure. Methodologically I work on ontology-grounded representations of reporting requirements, benchmark and diagnostic design, and reproducible research pipelines, alongside the panel econometrics and causal inference from my earlier accounting work.

I am currently working with Zherui Wang (UNSW CSE), who developed the underlying ESG disclosure corpus, on whether language-model judgements about corporate disclosure can be trusted as measurement. Item-by-item agreement with human reviewers is the usual test; the question we care about is whether that agreement still holds once individual judgements are aggregated into the scores and rankings that people actually rely on. To study it, we are improving the database and building an independently annotated benchmark, so that model output can be set against human judgement made without sight of it, and against how far human annotators agree with one another.

I am looking for collaborators interested in LLM evaluation, knowledge representation, retrieval systems, and corporate disclosure research. My goal is to make claims about machine-read disclosure checkable, grounding language-model systems in explicit reporting standards, traceable evidence, and benchmarks that measure what they claim to measure.