Tong Zhang

CoLaP: Continual Multimodal Prompt Learning

Manuscript under review

Continual adaptation of frozen vision backbones through semantic clustering of VLM-generated captions, contrastive alignment, and teacher-student distillation, retaining image-only inference. Improved ImageNet-O accuracy from 63.24% to 68.50% over L2P, with evaluation across five out-of-domain benchmarks.

Qualcomm Mentorship Programme

Qualcomm AI Research · Remote · June 2025 - Present

Generative AI safety research through a year-long mentorship collaboration, including DiSCO. Ongoing work investigates diffusion-based red teaming and defense through joint and disjoint trajectory analysis, using early denoising signals to detect malicious intent, adversarial edits, and concealed malicious patches.

Software Developer · CAMEL-AI

Open-Source Multi-Agent Framework · Remote · June - December 2023

Co-developed and deployed task-driven LLM multi-agent workflows with task planning, expert role assignment, role-playing dialogue, and shared environment context. Designed dependency-aware task decomposition and expert coordination, and improved memory and context handling through structured insight extraction.

King Abdullah University of Science and Technology

PhD Student · June 2025 - Present · Saudi Arabia

MS, Computer Science (with Thesis) · September 2023 - June 2025

Southern University of Science and Technology

Bachelor of Engineering, Computer Science · August 2019 - August 2023 · China

Final-year project: rubric-guided LLM report grading, comparing zero-shot and few-shot prompting against human scores.

I also proudly spent six months as a UN International University Volunteer (UNV) in Macedonia. Although it was a world away from my current research, it remains an amazing and meaningful experience.