Agentic Visual Generation: A Survey is now available on Preprints

Agentic Visual Generation: A Survey is now available as a public preprint.

PaperDOIGitHub8

Invited to serve as an Area Chair for ICLR 2027

FlowMM framework for information-flow-guided merging and sensitivity-adaptive token matchingPaper

FlowMM accepted to EMNLP 2026 Main

FlowMM adapts KV-cache merging to layer-wise cross-modal information flow and token sensitivity, reducing KV-cache memory by 80–95% and delivering 1.3–1.8× decoding speedup while preserving multimodal context performance.

PaperBlog

Invited to the AAAI 2027 Senior Program Committee

ACM Multimedia 2026 logoPaper

Three papers accepted to ACM MM 2026

VertiKV, PruneVAR, and SPECTRA are accepted to ACM MM 2026 — on vertical-integrity KV-cache compression for efficient multimodal long-context inference, training-free acceleration of visual autoregressive generation, and autonomous edge–cloud GUI grounding on device, respectively.

GUI-Anything sidecar observer for long Claude Code sessionsOpen Source

GUI-Anything is open-source — a flight recorder for long Claude Code sessions

We open-sourced GUI-Anything, a sidecar Flow Observer for Claude Code that watches live session streams, visualizes explorations as timelines and flowcharts, and brings local wiki context back on demand without interrupting the coding flow.

GitHub288Project Page
Overview of the proposed SPECTRAPaper

SPECTRA accepted to ACM MM 2026

SPECTRA lets lightweight on-device GUI agents autonomously decide when to request cloud assistance by stress testing visual cognition and analyzing latent trajectories. It retains 95.60% of cloud-only grounding performance at a 39.24% average request rate, without autoregressive decoding for request assessment.

BlogPaper
Overview of the VertiKV structured KV-cache compression frameworkPaper

VertiKV accepted to ACM MM 2026

VertiKV preserves vertical token-refinement paths during structured KV-cache compression, delivering 1.2–1.9× decoding speedup and an 80–95% reduction in KV-cache memory while maintaining multimodal long-context performance.

Blog
DeskCraft desktop GUI agent benchmark overviewOpen Source

DeskCraft is open-source — a benchmark for desktop GUI agents

DeskCraft puts desktop GUI agents to work on real professional workflows — 538 executable tasks spanning office, design, video, audio, 3D, and multi-app software, with human-in-the-loop collaboration. Code and paper are out now.

GitHub91Paper
paper2anything — one paper PDF turned into slides, a poster, a webpage, and social-media coversOpen Source

paper2anything is open-source — turn any paper PDF into slides, a poster, a webpage, or social posts

We open-sourced paper2anything, 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.

GitHub431Project Page
Repo-Docs skill for readable, understandable project codeOpen Source

Repo-Docs is open-source — help agents and humans understand your code

We open-sourced Repo-Docs, a coding-agent skill that turns fast-changing repos into something you can actually read. It builds and keeps repo-docs/ in sync with the code—walkthroughs, concept pages, glossary, change log, and agent rules—so newcomers can trace real behavior and understand design without hunting through files or chat.

GitHub467README