Programmable Spatial Intelligence:
AI can intelligently generate and perceive spatial worlds, reason and interact within them,
and verify outcomes using executable, symbolic, and semantically rich world representations.
Building Programmable Spatial Intelligence: AI can intelligently generate and perceive spatial worlds, reason and interact within them, and verify outcomes using executable, symbolic, and semantically rich world representations. This agenda grew out of building spatial agents to test spatial software after winning the National Software Testing Contest in a field of more than ten thousand participants. I now model objects, relations, physical rules, actions, and evolving states as a computational substrate for AI. This makes spatial worlds executable and checkable: actions update state according to physical rules, and agents can determine whether the intended outcome was actually achieved. My research brings formal methods, program analysis, and lifecycle assurance into XR (VR / AR / MR), embodied AI, and 3D world models.
CUHK PhD, advised by Prof. Michael R. Lyu (ACM / AAAS / IEEE Life Fellow). Currently on the founding team at Common Defense (Quantstamp). Previously co-founder & CTO of PortalX (backed by YC China · MiraclePlus), a GUI-agent startup; co-founder & CTO of InsightX, extending this agenda from virtual spaces into everyday physical-world context through personal world models and always-on wearable agents.
Across academic and industry R&D, I have served as technical lead and primary proposal author for joint-lab projects on large models and coding agents, taking work from grant applications and technical framing through system delivery. That engineering loop of structured representations, tool execution, testing, and feedback now informs my approach to Programmable Spatial Intelligence.
My academic service includes chairing workshops on XR software engineering and spatial intelligence, serving on the program committees of ICSE, FSE, ISSTA, IEEE VR, and ISMAR, and reviewing for ACM TOSEM, IEEE TSE, ACM Computing Surveys, among others.
From XR software reliability to Programmable Spatial Intelligence: building worlds that AI can read, write, act in, and verify.
From early WebXR reliability to software analysis that makes failures in spatial experience observable and testable.
Where spatial software reliability began.
The first systematic study of bugs in WebXR applications. It treated objects, user actions, state changes, and interaction consequences as software reliability problems that could be studied and checked.
Stereoscopic visual inconsistencies in VR.
Taxonomized the stereoscopic visual inconsistencies that cause cybersickness, then built StereoID, a static analyzer that surfaced them across 171 VR apps.
Tie intent to the right objects in 3D context, then turn those grounded actions into executable interaction.
Context-sensitive GUI grounding for XR testing.
A context-sensitive grounding approach that lets automated testers locate XR GUI elements the way human players do, by reading the surrounding 3D context rather than raw pixels.
Natural-language XR scripts.
Turns plain-English XR test instructions into executable controller-and-hand-tracking scripts. Closes the gap between human testers and the devices they wield.
Represent objects, relations, and constraints so a generated world can be modified and checked, not only rendered.
Constraint-expressive IR for LLMs.
An intermediate representation that lets LLMs reason about physical and semantic constraints when synthesizing 3D scenes. The resulting layouts are physically consistent, not just plausible.
Study how shared worlds diverge as people act, and what the system must preserve as state changes over time.
Demystifying multi-user XR defects.
First systematic study of defects unique to multi-user XR (desync, ghost avatars, shared-state corruption) across 100+ real-world apps.
Agent systems make representation, tool use, execution, and feedback operational; personal world models extend that loop beyond the headset.
Multi-modal software-onboarding agents.
Co-founded a startup backed by YC China and MiraclePlus, building universal self-serve AI assistants for any software. Multi-modal LLM agents do intelligent Q&A, automatic onboarding, and task automation across platforms.
Personal world models for always-on wearable agents.
Co-founded an AI startup turning fleeting physical-world conversations into persistent, callable context. Its lightweight wearable hardware combines multi-modal LLM agents with a personal world model to capture, structure, and act on real-world information.
Featured by ByteDance Volcano Engine at Force Link · early orders from Huawei & PwC
VR/AR Expo China
Internal invited talk at Huawei's 2012 Laboratories
Ranked Top 10 on the monthly trending list.
Technical Salon, CUHK-Huawei Foundation Models and Interactive Intelligence Innovation Laboratory
Huawei Workshop on Frontier Technologies for Ecosystem Testing
Technical Salon, CUHK-Huawei Large Model Reliability and Applied Engineering Innovation Laboratory
Graduate Student Forum hosted by the Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies
▸ first-author · # co-first-author · * corresponding
A context-sensitive grounding approach that lets automated testers locate XR GUI elements the way human players do, by reading the surrounding 3D context rather than raw pixels.
First systematic study of defects unique to multi-user XR (desync, ghost avatars, shared-state corruption) across 100+ real-world apps.
An intermediate representation that lets LLMs reason about physical and semantic constraints when synthesizing 3D scenes, producing layouts that are physically consistent and human-plausible.
RealityCraft turns plain-English XR test descriptions into executable interaction scripts, closing the gap between human testers and the controllers they use.
CARE mines user reviews of XR apps and uses an LLM pipeline to surface cybersickness symptoms developers can act on, turning subjective discomfort into reproducible signal.
Stereoscopic visual inconsistencies (SVIs) cause cybersickness in VR. We taxonomize them and build StereoID, a static analyzer that finds SVIs across 171 VR apps.
The first systematic look at bugs in WebXR applications, a foundational study for my Programmable Spatial Intelligence research agenda.
An audit of rendered-image evaluation across 300 frozen scenes from six text-to-3D generators. Camera settings and caption wording can change scores and point-estimate winners, motivating explicit evaluation protocols and selection-aware uncertainty when comparing generators.
MatLoom generates physically based materials as compact, executable layer programs. Shared spatial expressions make dependencies between coverage, color, and relief explicit, while parser-guided repair, preview-based critique, and seed search refine materials without task-specific fine-tuning.
The first large-scale empirical study modeling VR software quality from users' perspectives, mining 1.1M+ reviews of 14,150 VR apps across seven stores to build a taxonomy of 12 quality attributes that matter most to users.
Mentored 20+ students at the junior PhD, master's, and undergraduate levels across CUHK, HIT, SYSU, and SUSTech.
Guided students to publish at flagship venues, including ICSE and FSE.
Mentees have gone on to CMU, UW, EPFL, CUHK, HKU, HKUST, and NUS, as well as ByteDance, Tencent, and Alibaba.
I write novels. My favorite books include Gone with the Wind and the dragon book 🐉 Compilers: Principles, Techniques, and Tools.