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).

This research proposal focuses on the intersection of trustful artificial intelligence (AI) and the digital economy, aiming to develop frameworks and technologies that ensure AI systems are reliable, transparent, and beneficial to economic growth.

Hosted by the Singapore University of Technology and Design (SUTD), in collaboration with Singapore Management University (SMU), this project aims to create breakthroughs in wearable fabrics for health and wellness applications, with a special focus on tracking the mobility and joint motion of individuals with frailty challenges. The project aims to develop self-sustaining, energy-efficient smart textiles that seamlessly integrate sensors and electronics into knitted fabrics. The research collectively addresses two critical challenges: developing wearable fabric materials that can both sense movement and harvest energy, and creating ultra-low-power, on-board data processing mechanisms. The project combines all-knitted energy harvesting (using advanced yarns and knit architectures) techniques with ultra-low-power spiking neural network (SNN) based approaches for data processing, thereby maximising personal comfort while significantly extending the operational lifetime of the wearable sleeves. The research outputs will lay the foundations for scalable, long-term deployment of smart textile-based wearables for healthcare, rehabilitation, and preventive monitoring. Scientifically, the work shall generate globally competitive advances in materials science and neuromorphic computing for next-generation wearables; economically, it advances the marketability of textile-based wearables; and societally, it supports healthier ageing through quantified tracking of frailty and mobility-related impairments.

This research is supported by the National Research Foundation, Singapore under its 33rd Competitive Research Programme (NRF-CRP33-2025-0007).

Multimodal large language models (MLLMs) deliver strong generalization but suffer from severe inference bottlenecks: slow, costly, and energy-intensive decoding is increasingly driven by verbose explicit reasoning (e.g., chain-of-thought), which improves accuracy yet inflates token length and compute (“overthinking”). Latent reasoning offers a promising alternative by replacing long textual rationales with compact soft “thought” tokens that condition the target model, but current methods face two key barriers: (i) misalignment, since thought tokens are produced by a separately trained smaller assistant and do not match the target model’s internal representations, and (ii) poor cross-domain generalization, as a single assistant cannot cover diverse reasoning styles (math, code, dialogue, multimodal tasks).

We propose an efficient and scalable latent reasoning framework for MLLMs with two innovations. First, adapter-based thought token generation: a lightweight adapter transforms the target MLLM’s own intermediate features into thought tokens, improving alignment and preserving accuracy while reducing overhead. Second, domain-adaptive latent reasoning: a mixture of domain-specialized experts with a learned router selects the best expert per query to robustly support heterogeneous tasks. Together, these components aim to substantially accelerate MLLM inference while maintaining or improving reasoning quality.

Singapore’s strength lies in its cultural and linguistic diversity, where residents communicate daily across English, Chinese, Malay, Tamil, and local blends such as Singlish. While this enriches social resilience, it also creates communication challenges, especially as interactions with conversational AI become increasingly common. Current AI systems, trained mostly on English-based and Western-centric data, often fail to understand Singapore’s multilingual and multicultural expressions, limiting their effectiveness and reinforcing digital divides for communities such as elders. To address this gap, this project aims to develop proactive conversational AI agents that are culturally aware and linguistically flexible for the Singapore context. The research will proceed in three phases: (1) investigate communication breakdowns in multilingual human–AI interactions and build benchmark datasets for evaluating cultural conversational understanding; (2) enhance the cultural adaptability of large language models to ensure accurate, value-aligned responses across languages; and (3) deploy culturally adaptive agents in elder-care settings to maintain engaging, meaningful interactions. The project will deliver benchmark datasets, model audits, a culturally adaptive LLM backbone, and a proactive conversational agent, ensuring AI strengthens social resilience and supports diverse communities in Singapore and beyond.

VISTA is a research initiative that equips autonomous AI agents with an explicit and continuously tracked representation of human values, which enables safe and trustworthy decision-making during complex and long-horizon tasks. The architecture integrates real-time value monitoring, auditing and adaptive correction directly into the agent’s planning and optimization process, rather than relying on post-hoc safeguards. VISTA aims to support the deployment of high-impact autonomous systems that are both performance-efficient and aligned with emerging AI governance requirements.

This research/project is supported by the National Research Foundation, Singapore under its AI Singapore programme (AISG Award No: AISG3-RPGV-2025-017).

The project delivers an AI-Agentic Penetration Testing Platform that automates the full penetration testing workflow from reconnaissance to reporting, using large language models, agentic workflows, and interoperability with standard security tools. The platform will be codeveloped and piloted with Ensign InfoSecurity, ensuring alignment with industry standards and real-world needs. Initial deployment will focus on regulated sectors and enterprises in Singapore, with Managed Security Service Providers (MSSPs) as key distribution partners.

This project addresses the critical fragmentation in today’s software security landscape, where traditional vulnerability databases fail to keep pace with the velocity of modern threats. Led by Singapore Management University in collaboration with digiDations, this project develops an end-to-end, AI-agent-driven platform designed to autonomously transform noisy, multi-source signals into actionable, high-confidence security intelligence. The platform utilizes specialized autonomous agents to orchestrate the entire intelligence lifecycle through four core functions: adaptive discovery across over 20 heterogeneous sources (including informal channels), LLM-powered semantic normalization, automated conflict resolution with credibility scoring, and direct mapping to Software Bills of Materials (SBOMs). By systematically reconciling contradictory data and filtering misinformation, the system aims to significantly reduce operational noise and compress detection latency to under five minutes. The primary deliverable is a pilot-ready Minimum Viable Product (MVP) that creates a pathway to a commercial subscription offering. Ultimately, this project shifts security operations from reactive remediation to proactive prediction, building a sovereign capability that strengthens national cyber resilience against emerging software supply chain risks.