Qijian Tian

I'm a fourth-year Ph.D. student in Computer Science and Technology at Shanghai Jiao Tong University (SJTU), advised by Prof. Lizhuang Ma at the Digital Media & Computer Vision Laboratory (DMCV), with additional supervision from Dr. Xin Tan.

Before my Ph.D., I received my bachelor's degree in Computer Science and Technology from Beihang University (BUAA).

I also worked as an intern at Shanghai AI Laboratory and Baidu.

Email / Scholar / Github

Research

My research interests involve multimodal large language models (MLLM / VLM) and 3D vision.

I'm currently interested in VLM post-training and agentic RL, particularly in advancing the capabilities of foundation models for Computer-use Agents & Design Artifacts.

I have previously conducted some work in 2D/3D scene understanding, especially in scene parsing and 3D Gaussian splatting.

Report

Towards Safe and Trustworthy Embodied AI: Foundations, Status, and Prospects Embodied AI Safety Shanghai AI Laboratory
Technical Report 2025 project page / openreview

A framework for safe and trustworthy embodied AI, defining five levels and ten core principles, analyzing risks across the agent's operational stages, and outlining a roadmap toward safety across the closed-loop interaction between agents and their environments.

Paper
	From Views to Worlds: Active Exploration over 3D Worlds for Vision-Language Models VLM post-training Agentic RL Spatial Reasoning Qijian Tian, Jiayu Ying, Ke Fan, Lizhuang Ma, Xin Tan *NeurIPS 2026* project page / arXiv An active exploration framework that enables VLMs to actively explore 3D worlds and lifts VLMs from passive view-centric perception to active world-centric spatial reasoning.
	Exemplar2VQA: A Scalable Exemplar-Driven Visual Question Answering Generation Framework via Multi-Agent Coding VLM post-training Agentic Data Synthesis Spatial Reasoning Jiayu Ying, Qijian Tian, Ruijie Xu, Xinnan Zhu, Daoguo Dong, Jiachen Xu, Xin Tan *NeurIPS 2026* project page / arXiv An exemplar-driven framework that uses collaborative coding agents and geometric utilities to generate spatial question-answer pairs at scale in simulated 3D environments, improving VLMs' spatial reasoning through synthetic training data.
	FLEG: Feed-Forward Language Embedded Gaussian Splatting from Any Views via Compact Semantic Representation 3D Vision Open-Vocabulary Understanding Multi-view Input Qijian Tian, Jiayu Ying, Xuhong Wang, Yuan Xie, Lizhuang Ma, Xin Tan *ECCV 2026* project page / arXiv A feed-forward network that reconstructs language-embedded 3D Gaussians from arbitrary uncalibrated and unposed images with a compact semantic representation that avoids per-Gaussian language embedding and significantly reduces storage overhead.
	S2D: Sparse to Dense Lifting for 3D Reconstruction with Minimal Inputs 3D Vision Image Generation Yuzhou Ji, Qijian Tian, He Zhu, Xiaoqi Jiang, Guangzhi Cao, Lizhuang Ma, Yuan Xie, Xin Tan *CVPR 2026* project page / arXiv A sparse-to-dense lifting pipeline that bridges point clouds and 3D Gaussian Splatting via a one-step diffusion model, enabling high-quality 3DGS reconstruction with minimal inputs.
	DrivingForward: Feed-forward 3D Gaussian Splatting for Driving Scene Reconstruction from Flexible Surround-view Input 3D Vision Driving Scene Depth Estimation Qijian Tian, Xin Tan, Yuan Xie, Lizhuang Ma *AAAI 2025* project page / arXiv A feed-forward Gaussian Splatting model that reconstructs driving scenes from flexible sparse surround-view input.
	DANIM: Domain Adaptation Network with Intermediate Domain Masking for Night-time Scene Parsing Domain Adaptation Scene Parsing Qijian Tian, Sen Wang, Ran Yi, Zufeng Zhang, Bin Sheng, Xin Tan, Lizhuang Ma *Pattern Recognition 2026* A novel domain adaptation network for night-time scene parsing that bridges the day-night domain gap using an intermediate domain.
	UniForward: Unified 3D Scene and Semantic Field Reconstruction via Feed-Forward Gaussian Splatting from Only Sparse-View Images 3D Vision Open-Vocabulary Understanding Qijian Tian, Xin Tan, Jingyu Gong, Yuan Xie, Lizhuang Ma *Virtual Reality & Intelligent Hardware 2026* arXiv A feed-forward Gaussian splatting model that jointly reconstructs 3D scenes and semantic fields from sparse, uncalibrated and unposed images, enabling real-time rendering and open-vocabulary segmentation.
	Multimodal Place Recognition With Feed-Forward Gaussian and Mamba-Based Enhancement in Mixed Traffic Environments Place Recognition 3D Vision Jingyi Xu, Zhangshuo Qi, Xianyun Jiao, Qijian Tian, Zhongmiao Yan, Qi Wu, Xieyuanli Chen, Ling Pei *IEEE Transactions on Intelligent Transportation Systems 2026* A multimodal place recognition framework that combines feed-forward Gaussian representations with Mamba-based enhancement for mixed traffic environments.
	GroupForward: Building Referable 3D Scenes via Instance-Grouped Feed-Forward Gaussian Splatting 3D Vision Instance-level Understanding Referring Segmentation Qijian Tian, Zimeng Wu, Xuhong Wang, Lizhuang Ma, Xin Tan arXiv 2026 arXiv A feed-forward Gaussian splatting model that reconstructs instance-grouped 3D scenes from sparse, unposed images and supports complex referring segmentation through scene graphs and VLM-based reasoning.
	Calibrating Attention-Based Token Pruning for Efficient Multi-Image Synthesis in UMMs Image Generation Unified Multimodal Models Zimeng Wu, Qijian Tian, Penghui Du, Changqian Yu, Jiaxin Chen, Yunhong Wang arXiv 2026 arXiv A training-free token pruning framework that calibrates attention scores and synchronizes pruning across conditioning branches for efficient multi-image synthesis in unified multimodal models, achieving up to 2.09× attention-core speedup.
	DailyArt: Discovering Articulation from Single Static Images via Latent Dynamics Image Generation 3D Vision Articulated Objects Hang Zhang, Qijian Tian, Jingyu Gong, Daoguo Dong, Xuhong Wang, Yuan Xie, Xin Tan arXiv 2026 project page / arXiv A framework that reveals hidden articulation cues by synthesizing an opened state from a single static image, enabling joint estimation and part-level novel state synthesis.
	Generalized Category Discovery in Semantic Segmentation Scene Parsing Semantic Segmentation Zhengyuan Peng, Qijian Tian, JianQing Xu, Yizhang Jin, Xuequan Lu, Xin Tan, Yuan Xie, Lizhuang Ma arXiv 2023 arXiv A new task setting, Generalized Category Discovery in Semantic Segmentation (GCDSS), aims to discover and segment base and novel categories using labeled base-class data.
	LibEpidemic: An Open-source Framework for Modeling Infectious Disease with Bigdata Honghao Shi, Qijian Tian, Jingyuan Wang, Jiawei Cheng CIKM 2022 An open-source framework for infectious disease modeling with big data, enabling researchers to build subdivided, multi-stage and metapopulation models without programming.

This homepage's source code is from Jon Barron's website.

Research

Report

Paper