Ruiyu Mao

Ruiyu Mao

Ph.D. Student in Computer Science

The University of Texas at Dallas

I am a Ph.D. student in Computer Science at the University of Texas at Dallas (expected May 2027), advised by Prof. Yunhui Guo. My research focuses on Computer Vision & Data-Efficient Learning and Efficient LLMs & Multimodal Large Language Models. Specifically, I work on active learning, submodular optimization, data selection, 3D object detection, BEV perception, LiDAR perception, vision-language models, and multimodal DPO.

Richardson, TX, USA


News


Research Interests

Computer Vision & Data-Efficient Learning

Active learning, submodular optimization, data selection, open-set learning, 3D object detection, BEV perception, LiDAR perception, LiDAR semantic segmentation, autonomous driving.

Efficient LLMs & Multimodal Large Language Models

Vision-language models, multimodal learning, vision token reduction, attention-driven self-compression, efficient training, efficient inference, learnability-aware training, active data selection, multimodal DPO.


Publications

CVPR 2026

Learnability-Driven Submodular Optimization for Active Roadside BEV Perception

Ruiyu Mao, Baoming Zhang, Nicholas Ruozzi, Yunhui Guo

Proposed LH3D, a learnability-driven submodular framework for roadside monocular 3D detection that explicitly modeled ambiguous roadside-only scenes and selected camera views that were both highly informative and reliably labelable on DAIR-V2X-I.

LH3D Pipeline
BigData 2025

Vision Token Reduction via Attention-Driven Self-Compression for Efficient Multimodal Large Language Models

Omer Faruk Deniz, Ruiyu Mao, Ruochen Li, Yapeng Tian, Latifur Khan

Designed an attention-driven self-compression mechanism that leveraged the LVLM's own attention patterns to aggregate dense vision tokens into a compact set of summary tokens, removing redundant vision tokens while preserving accuracy.

ADSC Pipeline
NeurIPS 2024

STONE: A Submodular Optimization Framework for Active 3D Object Detection

Ruiyu Mao, Sarthak Kumar Maharana, Rishabh K. Iyer, Yunhui Guo

Proposed STONE, the first submodular optimization framework for active 3D object detection, with a two-stage algorithm achieving strong label efficiency on KITTI and Waymo.

STONE Pipeline
AAAI 2024

Inconsistency-Based Data-Centric Active Open-Set Annotation

Ruiyu Mao, Xu Ouyang, Yunhui Guo

Introduced NEAT, a data-centric active open-set annotation method that used CLIP features and label clusterability to detect known-class samples, and an inconsistency score to select informative queries while avoiding unknown classes.

NEAT Pipeline
Under Review

SELECT: A Submodular Approach for Active LiDAR Semantic Segmentation

Ruiyu Mao, Sarthak Kumar Maharana, Xulong Tang, Yunhui Guo

Developed SELECT, a voxel-centric submodular pipeline for LiDAR semantic segmentation on SemanticPOSS, SemanticKITTI, and nuScenes under limited annotation budgets.

SELECT Pipeline
In Prep

Learnability-Aware Active Multimodal Direct Preference Optimization

Ruiyu Mao, Omer Faruk Deniz

Developed an active multimodal DPO framework that used learnability-aware scoring to select informative image–text preference pairs, improving multimodal alignment while reducing preference data and training cost.


Education

The University of Texas at Dallas

Ph.D. in Computer Science · GPA: 3.62/4.00

Jan. 2024 – May 2027 (Expected)

The University of Texas at Dallas

M.S. in Computer Science · GPA: 3.62/4.00

Aug. 2021 – Dec. 2023

Case Western Reserve University

B.S. in Aerospace Engineering

Aug. 2017 – May 2021


Professional Service

Reviewer: CVPR, AAAI, NeurIPS, TMLR