Projects
Back to the Future: Regressing Readability Features from LLMs arXiv
BTTF combines linear regression with traditional code measurements, embedding-based semantic organisation, and causal-LM predictability to assess code readability. Evaluated against human judgements and controlled code transformations, it provides interpretable, feature-level explanations of readability.
When Specifications Conflict: A Symmetry-Based Framework for Measuring LLM Preferences arXiv
A symmetry-based framework for analysing how LLMs resolve competing task specifications, evaluated on 550 executable conflict instances across 11 mathematical function families and extended to Boolean algebra, code generation, and clinical-rule settings.
Cognitive Modeling of Code Readability: Working Memory and Semantic Recoverability MEng Thesis
My MEng thesis, supervised by Prof. Earl T. Barr, proposes CognaScore and PreFixScore: complementary cognitive models of code readability based on working-memory load and semantic recoverability, evaluated at both function and file level.
Meow-Omni 1: A Multimodal Large Language Model for Feline Ethology NeurIPS 2026
An open-source quad-modal large language model that integrates video, audio, physiological time-series, and textual reasoning for feline intent recognition.
HoarePrompt: Structural Reasoning About Program Correctness in Natural Language ICSE 2026 Distinguished Paper Award
A framework for verifying program correctness from natural-language specifications using Hoare triples, strongest postcondition calculus, and k-induction for reasoning about loops.
Axiom of Choice Presentation
Well-Ordering Theorem, Axiom of Choice, and Zorn’s Lemma, Applications of Zorn’s Lemma, Banach-Tarski Paradox, Independence of the Axiom of Choice
Completed during the summer after my first year, this was my first research experience. I am especially grateful to my friend and colleague Travor Liu for his help and support.