Yunhao Luo

🚀Excited to share our @NeurIPSConf paper Compositional Diffuser (CompDiffuser)! CompDiffuser scales planning horizons at test time, able to construct long-horizon plans while only trained on short-horizon data. Project page: comp-diffuser.github.io Thread 👇(1/n)

@zhiyuanjsu Thanks. As in Section 4.3, we propose multiple objectives to optimize for each factor x^i (i.e., short-horizon videos). By doing so, we enforce consistency between factors, and hence are able to use them to construct coherent long-horizon plans.

@yluo_y Interesting approach with CVP. How does the message-passing scheme specifically address the challenge of long-horizon planning beyond short-horizon diffusion models? THX!

Introducing Compositional Visual Planning (CVP)🤖✨#ICLR2026 How can we scale short-horizon video diffusion models for test-time long-horizon planning? We propose a completely training-free method!🤯 We frame the problem as inference over a chain-structured factor graph and use a novel message-passing scheme on clean Tweedie estimates to enforce strict boundary agreement. 🧩

Very cool work on generating coherent videos over pre-defined long camera trajectories. Glad to see our work CompDiffuser (comp-diffuser.github.io) played a part in inspiring this. Congrats to the team on the #ICLR2026 paper!

How can visual planning agents 𝙨𝙚𝙡𝙛-𝙞𝙢𝙥𝙧𝙤𝙫𝙚 from their own collected experience? We present 𝗦𝗜𝗟𝗩𝗥🩶, a framework that combines offline data with online experience for concurrent zero-shot generalization and sample-efficient self-improvement capabilities!#ICLR2026

Check out our #ICLR2026 oral paper CDGS! We show that modular generative models can be composed via inference-time classical search plus score composition. This enables new capabilities such as synthesizing large images from small generators and extending short-horizon video planners into long-horizon plans.

🚨 New paper alert !! 🎥 Video VLMs are strong at high-level semantics and long-range temporal understanding. 🧠 JEPA is almost the opposite: better at dense, high-frequency dynamics, local physical consistency, and fast corrective control, but are less suited for rich semantic reasoning and long-horizon reasoning. We try to get the best of both: 🧩 A VLM as a cortex-like reasoner for semantics and long-horizon planning ⚡ A JEPA branch as a cerebellum-like controller for fine-grained dynamics, physical consistency, and rapid corrections Proudly, we present ThinkJEPA: a VLM-guided latent world model that FiLM-fuse the pyramid repr of VLMs encoding long-horizon semantic reasoning into the JEPA repr for fine-grained, physically consistent dynamics prediction. 🔗 Project: zhanghaichao.xyz/ThinkJEPA/ 📄 Paper: arxiv.org/pdf/2603.22281

Introducing EgoVerse: an ecosystem for robot learning from egocentric human data. Built and tested by 4 research labs + 3 industry partners, EgoVerse enables both science and scaling 1300+ hrs, 240 scenes, 2000+ tasks, and growing Dataset design, findings, and ecosystem 🧵

To bring generalist intelligent robots to the real world, we have to overcome the data scarcity problem. At Rhoda, we are solving it by reformulating robot policies as video generation. Today, we introduce the Direct Video-Action Model (DVA)

Robot memory methods are growing fast, but systematic evaluation is largely lacking. 📉 Introducing RoboMME: a new benchmark for memory-augmented robotic manipulation! 🤖🧠 Featuring 16 tasks across temporal, spatial, object, and procedural memory 🔗 robomme.github.io

Video is a natural and scalable interface for robot planning. LVP demonstrates impressive in-the-wild planning and execution results, opening numerous research avenues for large-scale video models in robotics, such as real-time reactivity, memory, geometry, and precision.

I will be at #NeurIPS2025 at San Diego⛱️ next week and present our spotlight paper comp-diffuser.github.io. Feel free to dm/email if you want to chat!

🙏Huge thanks to my amazing collaborators: Utkarsh A. Mishra @utkarshm0410 Yilun Du @du_yilun and Danfei Xu @danfei_xu. This project wouldn’t have been possible without them. Our whole team will be at NeurIPS – feel free to reach out if you’d like to chat! (7/7)

📢For more, please check out our: Website: comp-diffuser.github.io paper: arxiv.org/abs/2503.05153 code: github.com/devinluo27/com… (6/n)

🚀Excited to share our @NeurIPSConf paper Compositional Diffuser (CompDiffuser)! CompDiffuser scales planning horizons at test time, able to construct long-horizon plans while only trained on short-horizon data. Project page: comp-diffuser.github.io Thread 👇(1/n)