Our #CVPR2025 paper is out!.Inspired by Neural Collapse, we show that OOD samples lie far from the origin and class weights—tying together many prior methods. Huge thanks to @YaoQin_UCSB, @liuziwei7, and @JingkangY for the collaboration & OpenOOD support! .

Great to be part of the Alumni in Industry Committee session at #ISIT2025—shared my journey into industry and learned a lot from others on similar paths.

Had a great time at #CVPR2025!.Grateful for the chance to present my poster and connect with so many amazing researchers！

Excited to be at #CVPR2025! Come visit our booth 👇.Also thrilled to share two papers—please stop by our posters!.📄 Detecting OOD through the Lens of Neural Collapse: 📄 Distilling Multi-modal LLMs for Autonomous Driving:

Excited to present our latest work on hallucination detection at the ITA workshop ( this Thursday! Looking forward to discussing the effect of noise injection in quantifying model uncertainty!.

This work is in collaboration with @YaoQin_UCSB. And thanks to @liuziwei7, @JingkangY, and Jingyang Zhang, our method is also available on:.- OpenOOD codebase: - Live OpenOOD leaderboard:

Distance to decision boundaries reflects model confidence. As models are more confident with ID, we hypothesize and validate that ID samples reside further from decision boundaries than OOD. Leveraging this, our detector achieves SOTA effectiveness with negligible latency.

In this work, we explore OOD detection from a novel perspective—decision boundaries. We efficiently measure sample distance to decision boundaries (illustrated below) using our proposed closed-form estimation.

Excited to share our #ICML2024 paper! Detect Out-of-Distribution (OOD) samples efficiently and effectively with our detector: OOD samples are closer to decision boundaries, In-Distribution (ID) samples are further away!.