I am an AI Researcher at Tencent Hy in Singapore.
My research focuses on multimodal reasoning and generation, with an emphasis on efficient learning and reliable behavior. My work has been published at NeurIPS, ICML, ICLR, CVPR, ICCV, and ECCV, and in TIP and TPAMI.
Previously, I was a Senior Research Scientist at TikTok / ByteDance. I received my PhD from SUTD in 2024, advised by Ngai-Man Cheung, and received the Outstanding Thesis Award (one of 3 theses selected university-wide).
Selected papers are grouped by research area; highlighted entries are representative work. Full list on Google Scholar.
† Corresponding author * Equal contribution
Thinking with images through adaptive visual coding. A multimodal model learns when and how to use tools, writes executable code, and checks the visual evidence it creates. Tech lead and corresponding author.
Token-efficient segmentation through generative language modeling. Compact textual mask representations reduce sequence length and speed up inference, without a separate mask decoder. Tech lead and corresponding author.
Multimodal models trained with interleaved distillation and reinforcement learning for reasoning and long-horizon agentic tasks.
My contribution: post-training trajectory synthesis and benchmark evaluation for long-horizon agentic multimodal understanding.
GenLIP pretrains the vision encoder as a language generator: the ViT reads image patches and learns to predict the accompanying text. This autoregressive objective aligns visual pretraining with LLMs, without a contrastive loss or a separate text decoder.
Training integrated into mlfoundations/open_clip, with initial encoder support in Hugging Face's timm.
A multimodal LLM plans and a diffusion transformer draws; alternating reinforcement learning trains both for visual generation across diverse scenarios.
Task-specific experts improve event modeling for dense video captioning, moment retrieval, and highlight detection.
AttackVLM: Black-box red teaming of multimodal foundation models, exposing cross-modal vulnerabilities through transferable visual attacks.
Integrated in MMDT (Microsoft Research) and MirrorCheck (NVIDIA).
Highlighted in The Alan Turing Institute's report on LLMs in finance (2024).
A practical recipe for watermarking diffusion models through training or fine-tuning, including hidden trigger prompts for text-to-image models.
Highlighted in European Parliament briefing (2023) and the International Scientific Report on the Safety of Advanced AI (AI Seoul Summit, Seoul, 2024).
Highlighted in NIST's (U.S. Department of Commerce) OpenMFC workshop (2023).
At TikTok, I developed LLM/MLLM workflows and agentic content understanding systems for media safety, frontier risk detection, and policy alignment. Deployed in products serving over 1K QPS and reaching over 1B users worldwide, these systems delivered a statistically significant 2% reduction in the platform-wide user complaint rate.
I also received the 2025 ByteDance Staff Award (Individual) for improved user experience.
AdAM uses adaptation-aware kernel modulation to select source knowledge relevant to a target domain, enabling parameter-efficient few-shot image generation.
Highlighted in SUTD's IMPact research digest.
Systematic diffusion explains when label smoothing helps or hurts knowledge distillation; low-temperature transfer preserves the benefits of a label-smoothed teacher.
Low-temp distillation with label-smoothed teachers integrated in HCL (NVIDIA Research).
Fine-tuning retains source knowledge that conflicts with a new domain. RICK identifies and prunes incompatible filters for few-shot image generation.
Oral presentation at the 6th Forum on Trust and Privacy for the Future Digital World, Singapore, 2023.
Born-again distillation improves cross-domain few-shot classification without additional parameters or training data.
A unified testbed reveals diversity degradation as the main failure of few-shot image generation. A mutual-information objective alleviates it.
Reviewer. NeurIPS, CVPR, TPAMI, TIP, TIFS, TNNLS, TASL, TMM, TCSVT, CVIU.
Teaching Assistant. 50.021 Artificial Intelligence and 50.035 Computer Vision at SUTD.