I am a master’s student at Beihang University, supervised by Prof. Si Liu. Before that, I received my bachelor’s degree in Software Engineering from Beihang University.
My research interests include Agentic AI, MLLMs, AIGC, and Embodied AI. Welcome for discussion and collaboration, feel free to drop me an email.
✉️ I’ll be on the job market in 2027. Happy to connect about opportunities!
🔥 News
- 2026.10: 🔥 Release OneSearch-VL, a unified multimodal deep research agent for images and videos.
- 2026.09: 🎉 Gen-Searcher is accepted to NeurIPS 2026.
- 2026.08: 🎉 OmniAlpha is accepted to ACM MM 2026 as an Oral presentation.
- 2026.08: 🎉 Video-MME-Logical is accepted to Findings of EMNLP 2026.
- 2026.06: 🔥 Release Video-MME-Logical, using visual coding to generate controllable videos at scale for training and benchmarking.
- 2026.06: 🎉 OpenSubject is accepted to ECCV 2026.
- 2026.03: 🔥 Release Gen-Searcher, a deep research agent that searches for knowledge and reference images to guide image generation.
- 2026.02: 🎉 Four papers are accepted to CVPR 2026 (3 Main, 1 Findings).
- 2026.01: 🎉 AerialVLA is accepted to AAAI 2026.
- 2025.12: 🔥 Release EditThinker, unlocking iterative reasoning for image editors.
- 2025.12: 🔥 Release OneThinker, an all-in-one reasoning model for image and video understanding.
- 2025.11: 🔥 Release Thinking-While-Generating, interleaving textual reasoning throughout visual generation!
- 2025.09: 🎉 Temporal-R1 is accepted to the NeurIPS@NextVid 2025 as an Oral presentation.
- 2025.07: 🔥 Release LLaVA-ST, with code, models, and ST-Align data and benchmarks for fine-grained spatial-temporal understanding.
- 2025.02: 🎉 LLaVA-ST and VCT are accepted to CVPR 2025.
- 2025.01: 🎉 Point Cluster is accepted to ICLR 2025, and AL-Ref-SAM2 is accepted to AAAI 2025.
- 2024.07: 🎉 Dynamic Prompting is accepted to ACM MM 2024.
📝 Publications
* Equal contribution · † Project lead · ‡ Corresponding author. Full list on Google Scholar.
Agentic AI
Tech Report OneSearch-VL: Unified Multimodal Deep Research Agent for Image and Video
Unifies single-image, multi-image, and video deep research in one agent.
NeurIPS 2026 Gen-Searcher: Reinforcing Agentic Search for Image Generation
First work to explore training a deep research agent for generation.
arXiv 2025 EditThinker: Unlocking Iterative Reasoning for Any Image Editor
Enables MLLMs to use image generators as tools for agentic iterative reasoning and self-refinement of generated results.
Multimodal Understanding & Reasoning
EMNLP 2026 Video-MME-Logical: A Controlled Diagnostic Benchmark for Video Temporal-Logical Reasoning
Uses visual coding to generate controllable videos at scale for training and benchmarking.
NeurIPS@NextVid 2025 Oral Reinforcement Learning Tuning for VideoLLMs: Reward Design and Data Efficiency
Advances video reasoning through data-efficient reinforcement learning across eight video understanding tasks.
CVPR 2025 LLaVA-ST: A Multimodal Large Language Model for Fine-Grained Spatial-Temporal Understanding
Unifies fine-grained spatial and temporal localization for multimodal video understanding.
CVPR 2026 OneThinker: All-in-one Reasoning Model for Image and Video
One reasoning model for ten image and video tasks, evaluated on 31 benchmarks.
Tech Report LongCat-Next: Lexicalizing Modalities as Discrete Tokens
Unifies text, vision, and audio understanding and generation through a shared discrete-token framework.
Tech Report Mini-Omni-Reasoner: Token-Level Thinking-in-Speaking in Large Speech Models
Enables speech models to reason while speaking, combining internal reasoning with continuous speech generation.
Visual Generation & Editing
arXiv 2026 RPiAE: A Representation-Pivoted Autoencoder Enhancing Both Image Generation and Editing
A visual tokenizer that strengthens both image generation and editing while preserving reconstruction fidelity.
CVPR 2026 Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation
First framework to interleave textual reasoning throughout image generation.
ACM MM 2026 · Oral OmniAlpha: Aligning Transparency-Aware Generation via Multi-Task Unified Reinforcement Learning
Unifies transparency-aware image generation and editing through reinforcement learning across multiple RGBA tasks.
ECCV 2026 OpenSubject: Leveraging Video-Derived Identity and Diversity Priors for Subject-driven Image Generation and Manipulation
Scales identity-consistent image generation and editing with 2.5 million training samples mined from videos.
arXiv 2026 Uni-Edit: Intelligent Editing Is A General Task For Unified Model Tuning
Establishes intelligent editing as a single training task that improves understanding, generation, and editing together.
arXiv 2025 AIA: Rethinking Architecture Decoupling Strategy In Unified Multimodal Model
Shows that aligning cross-modal attention improves understanding and generation without decoupling the model architecture.
Embodied AI
AAAI 2026 AerialVLA: A Vision-Language-Action Model for Aerial Navigation with Online Dialogue
An aerial navigation model that proactively asks questions to resolve uncertainty and reach its destination.
CVPR 2026 Parse, Search, and Confirmation: Training-Free Aerial Vision-and-Dialog Navigation with Chain-of-Thought Reasoning and Structured Spatial Memory
Training-free aerial navigation that matches or outperforms several fine-tuned methods through reasoning and spatial memory.
Show FigureHide Figure: Parse, Search, and Confirmation: Training-Free Aerial Vision-and-Dialog Navigation with Chain-of-Thought Reasoning and Structured Spatial Memory
arXiv 2026 From Instruction to Event: Sound-Triggered Mobile Manipulation
Moves mobile manipulation beyond explicit instructions, enabling robots to act on environmental sound events.
Show FigureHide Figure: From Instruction to Event: Sound-Triggered Mobile Manipulation
ICLR 2025 Point Cluster: A Compact Message Unit for Communication-Efficient Collaborative Perception
Compact point clusters reduce communication costs while preserving the geometry and semantics needed for collaborative perception.
Show FigureHide Figure: Point Cluster: A Compact Message Unit for Communication-Efficient Collaborative Perception
Visual Segmentation & Grounding
CVPR 2025 Revisiting Audio-Visual Segmentation with Vision-Centric Transformer
Vision-centric audio-visual segmentation that separates sounding objects in mixed audio while preserving precise object boundaries.
AAAI 2025 Unleashing the Temporal-Spatial Reasoning Capacity of GPT for Training-Free Audio and Language Referenced Video Object Segmentation
Training-free video segmentation guided by audio or language, with performance competitive with fully supervised methods.
ACM MM 2024 Dynamic Prompting of Frozen Text-to-Image Diffusion Models for Panoptic Narrative Grounding
Repurposes frozen text-to-image diffusion models for precise narrative grounding through dynamically adapted language prompts.
Show FigureHide Figure: Dynamic Prompting of Frozen Text-to-Image Diffusion Models for Panoptic Narrative Grounding
📖 Education
- Sep 2024 – Present, Master’s student, School of Artificial Intelligence, Beihang University.
- Sep 2019 – Jul 2024, Bachelor of Software Engineering, Beihang University.
📄 Academic Service
Conference reviewer: AAAI 2026, CVPR 2026, ICLR 2026, ICML 2026.