Research

My current work at NUS focuses on multimodal AI safety. It connects to my earlier research and industry experience in robot learning, trustworthy reasoning, and long-form video understanding at Delta, NTU, and Columbia. A recurring question is how multimodal systems remain reliable when evidence is incomplete, misleading, or adversarial.

Research Interests

  • Multimodal AI safety and adversarial robustness
  • Real-world robot learning and embodied AI
  • Trustworthy LLM and MLLM reasoning
  • Long-horizon multimodal understanding and evaluation
  • Tactile/VLA-style manipulation, world models, and failure analysis

Multimodal AI Safety

At NUS, I work with Prof. Hwee Tou Ng to study how multimodal AI can detect harmful video content and remain reliable when adversarial inputs attempt to manipulate its behavior.

This work examines adversarial attacks, including prompt injection and attacks on systems built around multimodal language models, and investigates defenses.

Real-World Robot Learning and Embodied AI

My recent industry work at Delta focused on robot-learning infrastructure, tactile and visual policy learning, VLA-style manipulation, rollout evaluation, and failure-driven improvement for real-world manipulation. My interests also include world models for embodied decision-making.

A recurring question in this line of work is how embodied systems can monitor uncertainty, gather missing evidence, and recover when contact-rich execution or navigation plans start to fail.

Representative work: CARe (CoRL 2024), AED (NeurIPS 2024), Affordance-Guided Base Placement (AAAI 2026), VLN-NF (ACL 2026).

Trustworthy LLM and MLLM Reasoning

I study how language and multimodal models reason under incomplete, misleading, or physically constrained evidence, with an emphasis on false premises, underspecified affordances, abstention, and failure analysis.

This direction centers on evidence-seeking behavior and constraint-aware reasoning in embodied and multimodal settings.

Representative work: Trope (EMNLP 2024), VLN-NF (ACL 2026), ADAPT (ACL 2026).

Long-Horizon Multimodal Understanding and Evaluation

Across movies and long-form multimodal narratives, I build datasets, benchmarks, and evaluation setups for reasoning that unfolds over extended time horizons.

This direction focuses on temporal coherence, narrative structure, evidence tracking, and evaluation protocols for long-context multimodal understanding.

Representative work: VLN-NF (ACL 2026), Trope (EMNLP 2024), MovieCORE (EMNLP 2025), HERMES (ICCV 2025).

Earlier Representative Work

Earlier projects that still connect to my current research directions include the following themes.

Multimodal Question Generation / QA: CaKE (CVPRW 2023), GPN (TCSVT 2021), SRCMSA (ICASSP 2020).

3D Vision: MonoDTR (CVPR 2022), S^3 (CVPR 2021), OCID-Ref (NAACL 2021).

Robotics and Embodied Grounding: GDN (CoRL 2020), SCAN (AAAI 2022), OCID-Ref (NAACL 2021).

Data-Efficient Learning: SCAN (AAAI 2022), Class-agnostic Few-shot Object Counting (WACV 2021), ReDAL (ICCV 2021).