Xin Lai 赖昕

Research Scientist

TikTok
ByteDance
Email: laixin1998@outlook[DOT]com

Biography [CV]

I am currently a researcher at TikTok, researching on Large Multimodal Models. Before that, I obtained my Ph.D degree in 2024 from the Chinese University of Hong Kong (CUHK), supervised by Prof. Jiaya Jia and Prof. Liwei Wang. I also received the Bachelor's Degree at Harbin Institute of Technology (HIT) in 2020.

My current research interest includes Multimodal LLMs, Long-term Reasoning for LLMs.

We are actively hiring research interns in vision-language fundation models. If you're interested, welcome to contact me by email with your resume.

Recent News

[2024/06] 🔥 We release Step-DPO, a simple, effective, and data-efficient way to significantly enhance the mathematical reasoning abilities of LLMs. Welcome to the GitHub Page! The resulting model achieves 70.8% and 94.0% accuracy on MATH and GSM8K, respectively, outperforming GPT-4-1106, Gemini-1.5-Pro, Claude-3-Opus!
[2023/08] 🔥 We release LISA that unlocks the reasoning segmentation capabilities for multi-modal LLMs! Welcome to try Online Demo. Over 1,500 GitHub Stars!
[2023/07] Mask-Attention-Free Transformer is accepted by ICCV 2023.
[2023/03] We release Sparse Transformer: fast, memory-efficient, and easy-to-use implementation for window-based 3D Transformer, well optimized by low-level CUDA code.
[2023/03] Spherical Transformer accepted by CVPR 2023: significantly enhance distant objects recognition for LiDAR points.
[2022/03] 🔥 Stratified Transformer accepted by CVPR 2022: a pioneering fully transformer-based 3D fundamental network.

Selected Publications [Google Scholar]

	Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xiangru Peng, Jiaya Jia arXiv pre-print [paper\|code] 70.8% and 94.0% accuracy on MATH and GSM8K, respectively! Outperforms GPT-4-1106, Gemini-1.5-Pro, Claude-3-Opus!
	LISA: Reasoning Segmentation via Large Language Model Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, Jiaya Jia CVPR, 2024 (Oral, 3.3% acceptance rate) [paper\|code] Over 1,500 GitHub Stars!
	Mask-Attention-Free Transformer for 3D Instance Segmentation Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, Jiaya Jia ICCV, 2023 [paper\|code]
	Spherical Transformer for LiDAR-based 3D Recognition Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, Jiaya Jia CVPR, 2023 [paper\|code]
	Stratified Transformer for 3D Point Cloud Segmentation Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, Jiaya Jia CVPR, 2022 [paper\|code] A pioneering fully transformer-based 3D fundamental network.
	DecoupleNet: Decoupled Network for Domain Adaptive Semantic Segmentation Xin Lai, Zhuotao Tian, Xiaogang Xu, Yingcong Chen, Shu Liu, Hengshuang Zhao, Liwei Wang, Jiaya Jia ECCV, 2022 [paper\|code]
	Semi-supervised Semantic Segmentation with Directional Context-aware Consistency Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, Jiaya Jia CVPR, 2021 [paper\|code]

Open-source Projects

Sparse Transformer

Xin Lai, Fanbin Lu, Yukang Chen

Open-source Library

[code]

Fast, memory-efficient, and easy-to-use implementation for window-based 3D Transformer, well optimized by low-level CUDA code.

Experiences

July 2023 - Jan. 2024, Tencent Co.,Ltd.
Internship, Mentor: Dr. Yanpei Cao

Dec. 2022 - May 2023, Microsoft Research Asia (MSRA)
Collaboration, Mentor: Dr. Yuhui Yuan

Honors and Awards

Outstanding Reviewer, CVPR, 2023
Postgraduate Scholarship, CUHK, 2020-2024
Outstanding Graduate, HIT, 2020
National Scholarship (Top 1%), China, 2018
Fung Scholarship, Hong Kong, 2018
Provincial Merit Student (Top 2%), China, 2018
People's Scholarship, HIT, 2016-2018

Professional Services

Conference Services:
International Conference on Learning Representations (ICLR’24)
Winter Conference on Applications of Computer Vision (WACV’24)
Conference and Workshop on Neural Information Processing Systems (NeurIPS’23)
IEEE Conference on Computer Vision and Pattern Recognition (CVPR’22,23)
IEEE International Conference on Computer Vision (ICCV’21,23)
European Conference on Computer Vision (ECCV’22)

Journal Reviews:
IEEE Transactions on Image Processing (TIP)
Pattern Recognition (PR)

Teaching

2021-2022	Fall	ENGG 1110: Problem Solving By Programming
2020-2021	Spring	CSCI 3251: Engineering Practicum
2020-2021	Fall	ENGG 1110: Problem Solving By Programming

	Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xiangru Peng, Jiaya Jia arXiv pre-print [paper\|code] 70.8% and 94.0% accuracy on MATH and GSM8K, respectively! Outperforms GPT-4-1106, Gemini-1.5-Pro, Claude-3-Opus!
	LISA: Reasoning Segmentation via Large Language Model Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, Jiaya Jia CVPR, 2024 (Oral, 3.3% acceptance rate) [paper\|code] Over 1,500 GitHub Stars!
	Mask-Attention-Free Transformer for 3D Instance Segmentation Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, Jiaya Jia ICCV, 2023 [paper\|code]
	Spherical Transformer for LiDAR-based 3D Recognition Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, Jiaya Jia CVPR, 2023 [paper\|code]
	Stratified Transformer for 3D Point Cloud Segmentation Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, Jiaya Jia CVPR, 2022 [paper\|code] A pioneering fully transformer-based 3D fundamental network.
	DecoupleNet: Decoupled Network for Domain Adaptive Semantic Segmentation Xin Lai, Zhuotao Tian, Xiaogang Xu, Yingcong Chen, Shu Liu, Hengshuang Zhao, Liwei Wang, Jiaya Jia ECCV, 2022 [paper\|code]
	Semi-supervised Semantic Segmentation with Directional Context-aware Consistency Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, Jiaya Jia CVPR, 2021 [paper\|code]

Xin Lai 赖 昕