Hi there! I am a senior undergraduate at Huazhong University of Science and Technology (B.S. in Intelligent Science and Technology, 2023 - 2027), from Taipei, Taiwan.
I study vision-centric agents that build and revise visual representations to understand, create, and act reliably.
I am currently a research intern at Tencent Hunyuan with Tianyu Pang, and at VisionX Lab, NYU Courant with Prof. Saining Xie, where I spent Fall 2025 as an exchange student. Previously I worked at OpenDILab, Shanghai AI Lab with Yazhe Niu, and at CAS-SIAT with Prof. Min Yang.
I am applying for CS PhD programs starting Fall 2027 and exploring full-time research opportunities at frontier AI labs worldwide. I also welcome long-term research internships and collaborations. Please feel free to connect with me!
I work on the gap between what a model sees and what it understands: implication, spatial relationships, and how states change over time. I believe strong visual representations are fundamental to reasoning, generation, and action.
My focus is Agentic Vision & Multimodal Coding, with coding agents as the primary mechanism. My bet is that executable visual environments can provide vision with what code offers language models: a training ground where agents work with real scenes over long horizons, verify outcomes, and learn transferable skills. A vision agent gathers evidence, uses visual tools, and maintains and revises visual state. Through coding and inverse graphics, it turns that understanding into executable scene programs, then renders, inspects, and refines the results. This unifies visual understanding and generation within a single agent.
I believe inspectable states and verifiable feedback offer a path to scaling both learning and inference. The challenge is to connect learned representations, predictive world models, and agent systems so that improvements generalize across tasks and environments.
Long-horizon intelligence must also account for human goals. I am interested in emotion as an experience-shaped value signal that guides attention, memory, and decisions. My goal is to build vision-centric agents that see, remember, reason, create, and act reliably in the real world.
Publications, research systems, and open-source projects.
All research works and open-source projects.
Supports long-horizon visual reasoning and action by letting VLMs inspect images, use typed visual tools, retain task memory, and refine decisions through verifiable feedback.
ARC-AGI-3: 25/25 games · 183/183 levels · score 100.0 · 7,340 official actions (far fewer than human 17,135)
Turns long-form video creation into an executable workflow: plan shots from scripts, generate and visually review clips, repair selected shots, and assemble the final film.
Core Contributor
Trains multimodal models to interpret visual metaphors through end-to-end reinforcement learning, using verifiable true/false judgments to connect visual evidence with cultural, emotional, and contextual meaning.
Under Review
Interprets visual metaphors by combining grounded image descriptions with adaptively retrieved background knowledge and explicit reasoning, addressing the contextual information missing from direct image understanding.
Under Review
Evaluates how multimodal models interpret Chinese images, exposing gaps in cultural understanding through manually reviewed visual content, including traditional artworks, and human-written answers.
Evaluates implicit image understanding beyond object recognition, revealing weaknesses in abstract visual reasoning and showing how emotion cues affect models' interpretations.
Reconstructs multi-turn counseling dialogues from professional reports, then evaluates how language models apply counseling knowledge in context through turn-level assessments of response quality.
Separates psychological knowledge from case analysis when evaluating language models, using Chinese examination questions to test both factual understanding and its application to realistic scenarios.
Builds graph reasoning in language models with generated graph problems and intermediate solution steps, supporting both capability evaluation and instruction tuning on graph tasks.
Large language models for mental health. Built the end-to-end framework spanning data construction, post-training, evaluation and deployment. Post-trained 20+ LLMs (7B–32B) with SFT, OPD, and RL (DPO, PPO, GRPO, etc.), and scaled it to one of China's largest LLM-for-mental-health communities (15k+ users).
A community exploring the implementation path and future development direction of Sora.
The first decision intelligence platform covering the most comprehensive RL algorithms and applications in academia and industry. Contributed to LightRFT — RLHF/RLVR fine-tuning for LLMs and VLMs over vLLM and SGLang backends with FSDP2/ZeRO sharding — and to the core frameworks DI-engine and LightZero.
A collection of LLM papers, blogs, and projects focused on OpenAI o1 and reasoning techniques. Led the literature review on LLM reasoning and thinking with images.
A collection of research papers on Reinforcement Learning with Human Feedback (RLHF), one of four curated surveys I maintain on RLHF, RLVR, UI/GUI agents and exploration in RL.
A project on large language models (LLMs) and human social skills.
A community of high-quality LLMs with a full-stack toolchain for development and application.
A flexible, efficient and production-ready RL training library for large language models (LLMs).
An open-source PyTorch toolbox for image and video editing and generation.
No works match this topic. Choose another topic.
Wuhan, China
Supervised by Tianyu Pang
Focus: Multimodal RL post-training, Long-horizon visual agent, Multimodal evaluation, Video generation.
Supervised by Prof. Saining Xie
Focus: Thinking with images, Spatial intelligence, Explicit visual states, Visual state tracking and memory, Multi-step visual inference.
Supervised by Yazhe Niu, Chaochao Lu and Prof. Hongsheng Li
Focus: Multimodal RL post-training, Agentic RL for long-horizon trajectories, Latent world models, Open-source RL infrastructure.
Supervised by Minghuan Tan, Prof. Shiwen Ni and Prof. Min Yang
Focus: Multimodal benchmarks, Image implication understanding, Human-centered evaluation of language models.
Supervised by Prof. Hong Huang and Prof. Hai Jin
Focus: Graph understanding and reasoning, Instruction tuning for large language models.
