Byungwoo Jeon

Ph.D. student @kaist_ai

[장수혁 @gloryhyeok24]

A recent line of WAMs generates actions from VDM hidden states of a single denoising step, yet their quality gets little attention.

👉 We introduce FRED to improve them via self-distillation, with large gains on real robot experiments.

🌐 https://t.co/tE4SdJgMEX
🧵 [1/7]

[Sadao Tokuyama @tokufxug]

複数カメラの映像で物体の動きを一致させる「Video-Point Model（VPM）」🤖

🔹NVIDIA Cosmos 3を基盤に開発
🔹3D空間での動きを学習し、シミュレーション映像の品質指標FVDを14〜19%改善
🔹ロボットのゼロショット操作も実証

開発者：Sunghwan Hong氏ら
論文・データ・コード公開中。詳細はリプ欄。

[Byungwoo Jeon @rootyjeon]

@XinchenYinYXC Yes, we also trained in the monocular camera setting. In practice, most robotics and autonomous driving systems rely on multiple cameras. We believe extending to multi-view is an essential step toward the next paradigm of world models.

[Byungwoo Jeon @rootyjeon]

Can explicitly learning future point motion improve the model that generates those future images?

We introduce VPM (Video-Point Model), a 4D world model that jointly forecasts future multi-view videos + persistent 3D point trajectories.

Short answer: yes. Teaching a video world model to forecast 3D point tracks makes its videos more faithful & 3D-consistent, and makes robots better.

🌐 Project: https://t.co/srVtzZsY5c
🧵 [1/6]

[Byungwoo Jeon @rootyjeon]

Thanks! Indeed, 3D geometry cannot be fully recovered from monocular input alone. Even with depth, only a single 2.5D view is available, and occluded regions remain unknown. Multi-view input gives the model access to richer and more diverse observations than a single egocentric view. We discuss this empirically in Table 22.

[Byungwoo Jeon @rootyjeon]

📄 arXiv: https://t.co/FgWuwGJiUH
💻 Code: https://t.co/IKbFnUY7Vc
🤗 HF: https://t.co/0oNVAcIvP6

Co-led w/ Sunghwan Hong. Big-shout out to my advisor, Jinwoo Shin @JinwooShin0417. This project would not have been possible without my amazing collaborators Dohyeon Kim, Jonghoon Lee, Jungwoo Park, Inhee Lee, Doosan Baek, Jinil Kim, Taehyun Lim, Hanbyul Joo @jhugestar, Marc Pollefeys @mapo1, and Saining Xie @sainingxie.

KAIST · ETH Zürich · SNU · NYU · RLWRLD · Physics Sim Lab
🧵 [6/6]

[Byungwoo Jeon @rootyjeon]

Robots too. Add actions as a token modality and the point-supervised backbone gives the best success rates on LIBERO, RoboCasa and a real Franka arm: the tracks describe the robot's own future motion, densely and in 3D. Point track is highly correlated with robot actions.
🧵 [5/6]

[Byungjun Yoon @happyhappy_jun]

🏆1st place in the AI Scientist track at Ralphthon @ ICML2026 — $10,000 in OpenAI credits.
@happyhappy_jun & @WoominSong from @KAIST_AI

We've open-sourced all of it: harness, code, the full log of research traces.

See our post: https://t.co/HIzZ7dOELn

How we did it 🧵

[Byungwoo Jeon @rootyjeon]

@de_Jiung huge congrats!

[DailyPapers @HuggingPapers]

SpatialBoost enhances vision encoders with 3D spatial awareness

A framework that injects 3D spatial knowledge into pre-trained vision encoders via language-guided reasoning, boosting DINOv3 performance by 3.8% on ADE20K and advancing 3D perception tasks.