Keming Wu

PhD student @Tsinghua_Uni. Focus on Generative AI and VLM. Author of EditReward, OpenMMReasoner.

[Keming Wu @Keming_Charles]

Why do open-source image editing models lag behind closed-source giants like GPT-Image-1, Seedream, & Google-Nano-Banana? 🤔

It’s mainly due to the quality of the training reward signal.

We’re bridging the gap. Meet EditReward! 🏆

[Zuhao Yang @mwxely464]

What if the very pretrained prior that lets an RL agent explore tools also destroys the format that made it tool-native?

We name this the Tool Prior Paradox — and tame it with PARA-GRPO.

🚀 Introducing ParaVT: parallel video tool use × agentic RL.

[Keming Wu @Keming_Charles]

🙏 If you find this useful:

⭐ Star the repo → https://t.co/Mgc3urfLGg

👍 Upvote on HF Daily Paper → https://t.co/1Hn3mrdYDS

🔁 Retweet to help us reach researchers working on world models, video gen & reward modeling

[Keming Wu @Keming_Charles]

🌍 Can today's video generators REASON about how the world should evolve — or do they just render it beautifully?
Introducing WorldReasonBench: a human-aligned stress test that re-frames video generation as future world-state prediction.
🌐 https://t.co/c0HBr58WRu

[Keming Wu @Keming_Charles]

Qualitative example — when SOTA still fails 🎬
Visually plausible ≠ world-aware.
Even top models get classic prompts wrong:
• "Pencil in water" — refraction direction inverted
Browse all qualitative cases on the project page 👇

[Computer Vision and Pattern Recognition Papers @CSVisionPapers]

Visual Generation in the New Era: An Evolution from Atomic Mapping to Agentic World Modeling

Keming Wu, Zuhao Yang, Kaichen Zhang, Shizun Wang, Haowei Zhu, Sicong Leng, Zhongyu Yang, Qijie Wang, …
https://t.co/3PaATxbfiA [𝚌𝚜.𝙲𝚅]
💬Project: https://t.co/e06OsM7VfL

[Kaichen Zhang @KaichenZhang358]

Excited to share a fun project I recently collaborated on: a roadmap for thinking about where visual generation is heading next. The key question is no longer just “can it make beautiful images?”, but whether it can handle memory, interaction, and eventually world modeling.

[Keming Wu @Keming_Charles]

Takeaway:

The future is not just higher-fidelity images.

It is controllable, interactive, verifiable, and world-aware generation.
arXiv: https://t.co/F40gaRS939
HF Daily Paper: https://t.co/jnC75XxbGf
GitHub: https://t.co/G50LYzKrLJ
WebPage: https://t.co/cB510BFxfp

[Keming Wu @Keming_Charles]

Excited to share our new roadmap:

Visual Generation in the New Era: An Evolution from Atomic Mapping to Agentic World Modeling

What does it mean for image generation models to become truly intelligent?

HF Daily Paper: https://t.co/jnC75XxbGf
WebPage: https://t.co/cB510BFxfp

[Keming Wu @Keming_Charles]

Benchmarks often reward visual quality.

But real progress also needs spatial reasoning, topology, symbolic structure, and code/math-grounded correctness.

We stress-test physical and causal reasoning:
These examples probe the boundary between image synthesis and world modeling.

[Keming Wu @Keming_Charles]

An insightful and excellent piece of work.

[Lianghui Zhu @lianghui_zhu]

For a decade, we've made models wider and deeper—but we've barely changed how layers *talk* to each other.

Since ResNet's `x + F(x)` in 2015, the depth residual has been the only highway for inter-layer communication.

It's time to upgrade the staircase. 🧵

[Jianyang Gao @gaoj0017]

The TurboQuant paper (ICLR 2026) contains serious issues in how it describes RaBitQ, including incorrect technical claims and misleading theory/experiment comparisons.

We flagged these issues to the authors before submission. They acknowledged them, but chose not to fix them. The paper was later accepted and widely promoted by Google, reaching tens of millions of views.

We’re speaking up now because once a misleading narrative spreads, it becomes much harder to correct. We’ve written a public comment on openreview (https://t.co/nDVjmNhATM).

We would greatly appreciate your attention and help in sharing it.