PhD Student · Computer Science · Purdue University
— DripNowhy
Hi there! I'm a PhD student at Purdue University 
, Department of Computer Science, advised by Dr. Ruqi Zhang. I obtained my B.S. degree at the School of Mathematics, Tianjin University 
. Previously, I worked as a research assistant in the MLDM Lab's Multimodal Vision Processing (MVP) Group, under the guidance of Dr. Bing Cao.
Reinforcement Learning, On-Policy Distillation, Reasoning, Self-Correction, and Self-Evolving.
Preference Alignment, Safety Alignment, Reliable Supervision, and Robustness.
Open to collaborations! Feel free to reach out if our research interests align.
Our paper on on-policy self-adaptation (OPSA) was released.
Two papers were accepted to ICML 2026.
One paper was accepted by ICLR 2026.
One paper was accepted by NeurIPS 2025.
One paper was accepted by EMNLP 2025 Main Conference.
Yi will give a talk about VLM safety at Shenlan School.
Yi serves as reviewer for NeurIPS 2025.
Our paper, dataset, and models about VLM Multi-Image Safety (MIS) are released.
Our paper about MLLM safety alignment was accepted at ICLR 2025.
Yi serves as reviewer for ICLR 2025.
Our paper about dynamic image fusion without additional training was accepted at NeurIPS 2024.
Yi will present a poster at ICML 2024, Hall C 4-9 #2817, Vienna, Austria.
Our paper about multimodal fusion was accepted at ICML 2024.
We show that on-policy distillation improves largely by suppressing low-probability student tokens rather than transferring teacher knowledge, and introduce OPSA, a supervision-free method that turns token uncertainty into self-improvement.
We propose Entrocraft, a rejection-sampling method that precisely controls entropy schedules during LLM reinforcement learning, alleviating performance saturation while improving generalization, output diversity, and long-term training.
We propose Octopus, an RL rollout augmentation framework that synthesizes dense self-correction examples by recombining existing rollouts. Octopus-8B achieves SoTA performance by advancing reasoning and self-correction capabilities.
We propose modular safety guardrails with monitoring and intervention layers, and show how cross-layer co-design enables faster, less conservative, and more effective safety for physical AI.
We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety.
Introducing the first multi-image safety (MIS) dataset, which includes both training and test splits. The VLMs fine-tuned with the MIRage method and MIS training set improve both the safety and general performance of the models.
Loading repositories…
