External Research Grants

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.

This project builds upon algorithms previously developed by the Prof Lau Hoong Chuin under the Quantum Engineering Programme 2.0 initiative - specifically in addressing variants of the News Vendor and Knapsack Problems - to tackle the increasing complexity of consumer demand and fluctuating market dynamics in logistics. In collaboration with ST Logistics (STL), the project will develop a hybrid quantum-classical model capable of jointly performing demand forecasting and inventory optimization. The goal is to deliver a proof-of-concept (POC) solution with computational efficiency for complex, real-world logistics scenarios provided by STL.

The overall objective of this project is to develop systematic and rigorous ways of safeguarding foundation models, including large models such as large language models (LLMs) as well as large multi-modal models (LMMs), against state-of-the-art and future security attacks. While there have been many bandage-like mitigation approaches on mitigating security attacks on LMs, they are far from having a lasting effect. The reason is that these mitigation approaches are treating the symptoms rather than fixing the causes of the problems. The team aims to develop techniques and systems which can detect and defeat a variety of security attacks on large models, through either prompting, finetuning or instruction- tuning, for the goal of jailbreaking, or embedding backdoors. This research / project is supported by the National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative.

Instructors today face increasing challenges in designing and delivering courses that effectively balance cognitive load, align with intended learning outcomes, and actively engage diverse learners. Traditional lecture slides and assessments often lack structure, personalization, and interactivity leading to passive learning, reduced motivation, and inconsistent achievement of educational goals. Furthermore, evaluating and improving teaching delivery remains largely subjective, with limited tools to analyze real-time classroom engagement or instructional clarity. This project offers AI-driven analysis and nudging mechanisms that align content with learning objectives through semantic and topical modeling, while embedding cognitive design strategies to manage learners’ mental load effectively. This research addresses these challenges by exploring how AI and analytics can enhance course content design, assessment, and delivery in a data-informed and scalable manner.

Eye tracking has emerged as a powerful, non-invasive window into neurological and ocular health, offering early biomarkers for conditions such as Parkinson’s disease, Alzheimer’s disease, and glaucoma. However, current RGB camera–based systems are bulky, power-intensive, and limited in their ability to capture the subtle, high-frequency micro-movements of the pupil that are critical for early diagnosis. To overcome these limitations, this project introduces SynapSee, a novel end-to-end wearable system that integrates event cameras with a multi-light active probing setup and computationally optimised algorithms for real-time, fine-grained pupil tracking. Unlike conventional eye trackers, event cameras operate at sub-microsecond latencies and asynchronously capture changes in light intensity, making them uniquely suited for high-velocity saccades and micro-movements. By exploiting “dark” and “bright” pupil effects through multi-light probing, SynapSee reduces extraneous event volume, enabling low-power and efficient processing. The system is further enhanced by hybrid spiking neural networks, adaptive sensing algorithms, and collaborative offloading to nearby devices, achieving both accuracy and energy efficiency. We will validate SynapSee in two exemplar clinical contexts: (i) detecting early neurodegenerative changes in Parkinson’s disease and (ii) identifying the onset of low-vision conditions such as macular degeneration, cataracts, and glaucoma. Longitudinal user and patient studies, conducted in collaboration with clinical partners, will establish discriminative ocular biomarkers and benchmark the system’s sensitivity and specificity. By enabling unobtrusive, continuous, and large-scale monitoring via smart glasses, SynapSee has the potential to transform preventive healthcare, offering clinicians powerful tools for early intervention and personalised disease management.

This project focuses on creating self-adaptive embodied agents capable of perceiving and planning in dynamic real-world environments, addressing current challenges like hallucinated plans, poor object tracking and inflexible execution. It employs retrieval-augmented planning, fine-grained environment understanding, and adaptive plan refinement using large multimodal models, validated through simulations and real robots in household tasks. Expected outcomes include new methods for adaptive planning and perception, a kitchen activity video dataset, and demonstrations in domestic scenarios, with broad applications in autonomous vehicles and assistive devices. The initiative aims to impact daily living and healthcare, especially eldercare in Singapore, aligning with national priorities to enhance AI leadership and support the Smart Nation agenda.