Projects

A public survey on agentic visual generation, with the full manuscript hosted on Preprints and the source repository kept in sync with the lab website.

A sidecar Flow Observer for long Claude Code sessions. Claude Code stays native in the left pane while the right pane maps session streams into timelines, flowcharts, summaries, and reusable local wiki memory.

An agent skills pack that turns an academic paper PDF into a presentation deck, conference poster, project homepage, Xiaohongshu post, or WeChat article — each generated end-to-end and ready to share.

A coding-agent skill that makes fast-moving repos easier to read and understand. It writes and syncs repo-docs/ with the codebase—walkthroughs, concept pages, glossary, change log, and agent rules—so readers can follow one real path through the code and grasp how it works.

An open-source AI video production assistant that turns general-purpose agents into video production experts, with brief clarification, proposal gates, asset generation, composition, and review before expensive generation runs.

A desktop GUI benchmark for professional workflows and human-in-the-loop collaboration, with 538 executable tasks across office, design, video, audio, 3D, and multi-app software.

A family of compact GUI agents for computer use, evolving across three works: InfiGUIAgent — a multimodal generalist agent with native reasoning and reflection; InfiGUI-R1 — from reactive actors to deliberative reasoners via reinforcement learning; InfiGUI-G1 — stronger GUI grounding through adaptive exploration policy optimization.