I am Yunqing Zhao, I work on multimodal models / agents.
My research asks how models learn from limited data, how they hold up against untrusted or adversarial inputs, and how they use tools to gather evidence before they answer.
-
CodeDance — Think with images through code, with adaptive tool use. CVPR
-
GenLIP — Generative multimodal pre-training. ECCV
-
Text4Seg++ — Vision tasks recast as generative language modeling. TPAMI
-
AdAM — Adaptive parameter-efficient generative fine-tuning. NeurIPS
-
AttackVLM — Frontier risk evaluation for VLMs with pixel-only, black-box targeted attack. NeurIPS
-
ThinkGen — Reasoning for visual generation. CVPR
-
Pistis Tech report — Multimodal foundation and reasoning models, agents and long-horizon tasks. Tech Report