Yihong Sun

📢 Excited to share our latest #NeurIPS2025 work on tracking and understanding object transformations! 🔗: tubelet-graph.github.io Come chat with us on Friday 12/5, 11AM-2PM, at Poster #4703!

Excited to announce our #AISTATS📜on training LLMs with MXFP4! We use stoch. rounding and random Hadamard transforms (all fast on HW) to get low-variance, unbiased gradient estimates with MXFP4 GEMMs. We get a ~30% speedup over FP8 with almost no PPL gap! arxiv.org/abs/2502.20586

📢 Excited to share our latest work: Video Creation by Demonstration! arXiv: arxiv.org/abs/2412.09551 Check out our results here: delta-diffusion.github.io

Happy to get feedback and questions! For more details, please check out our paper! See you on Friday 10/4 at Poster Session 7, #73! 😃 Huge thanks to my advisor, @BharathHarihar3! This wouldn’t have been possible without your support! (7/n)

Results: We obtain significant improvements over previous unsupervised object detection methods across multiple datasets & metrics, with notable improvements in Box AR by 6.6 on Waymo Open, 5.9 on nuScenes, and 8.2 on KITTI compared to CutLER. (6/n)

📢Excited to share our latest @eccvconf work MOD-UV on learning object detectors from unlabeled videos only! 🔗: mod-uv.github.io I will give a talk at Wild3D workshop (cc: @weichiuma) on Monday 9/30 and come chat with us on Friday 10/4 at Poster Session 7, #73!

Thanks @arankomatsuzaki for sharing our #CVPR2024 ViTamin work! Collaborated with @yucornetto1, Xiaohui Shen, @YuilleAlan and Liang-Chieh Chen TLDR: We design a scalable vision model in the vision-language era, advancing the limits for VLMs and multi-modal LLMs!

Excited to share our work, ObjectCarver! Given multiview images and click points on one image, ObjectCarver decomposes scenes into separate objects, providing high-quality 3D surfaces while handling occlusion and close-contact objects. (1/6) website: objectcarver.github.io