External Research Grants

PerFormRect aims to detect the misconceptions of students and utilize these insights in a feedback loop for crafting personalized formative assessment questions in programming education using Large Language Models (LLM). This project addresses Theme 4 (Leveraging Technology to Enhance and Personalize the Learning Experience) with efficient analysis of individual coding misconceptions and timely intervention through personalized coding questions for both practice and self-assessment. It tackles the pressing need to efficiently identify coding misconceptions of students and provide timely, personalized questions for both practice and self-assessment by students – tasks which are difficult for human instructors to perform at scale. PerformRect leverages on LLM to identify students’ misconceptions and generate personalized feedback (WP1) based on their code submissions. The identified misconceptions will also be used to generate formative assessment coding questions (WP2) tailored for their repetitive learning. The effectiveness of PerFormRect on students’ learning will be evaluated using randomized control trials (RCTs) in two institutions (for generalizability). In all, PerFormRect offers on-demand access, personalized feedback, tailored assessments and scalable programming skills development opportunities, allowing for a broader reach without sacrificing quality.

This project empowers educators to effectively leverage Generative AI as a transformative tool that not only identifies and addresses student misconceptions but also deepens understanding, fosters critical thinking, and enriches the overall learning experience. It proposes a novel AI-enhanced pedagogical approach called Debunkr, which utilises a cognitive conflict instructional approach to actively debunk misconceptions for university courses, potentially adaptable for any subject in the age of Generative AI.

TITAN 2.0 is a 2-year project to build an advanced AI-driven framework that identifies and fixes complex security flaws in software code. By combining traditional static analysis with the reasoning power of large language models (LLMs) orchestrated in an agentic fashion, the system extends analysis beyond individual functions to comprehend complex interactions across multiple files and modules. This allows it to catch interprocedural vulnerabilities that simpler LLM-powered tools often miss. Designed to support multiple programming languages such as Java, C#, Python, and JavaScript, the framework does not just flag risks; it acts as a digital security partner by providing automated CWE labeling, validated code patches, and developer-friendly reports. By integrating these smart agents, the project significantly improves vulnerability remediation to ensure that digital services remain secure and resilient.

This research / project is supported by the National Research Foundation, Singapore, and Ministry of Digital Development & Information under its Smart Nation and Digital Government Translational R&D Grant (Award No: TRANS2026-TGC01).