Test-time scaling nailed code & math—next stop: the real 3D world. 🌍 . MindJourney pairs any VLM with a video-diffusion World Model, letting it explore an imagined scene before answering. One frame becomes a tour—and the tour leads to new SOTA in spatial reasoning. 🚀. 🧵1/

Just paid ¥4.99 to a site that "predicts" NeurIPS acceptance from your ratings and confidence scores. Total scam-basically a random number generator. 🤡.I should build my own startup for this. Pretty sure I could make a fortune off researchers' anxiety these days. #NeurIPS2025

RT @jw2yang4ai: VLM struggles badly to interpret 3D from 2D observations, but what if it has a good mental model about the world? . Checkou….

RT @gan_chuang: Spatial reasoning from a single image is inherently difficult, but it becomes significantly easier when leveraging a contro….

See our project webpage, paper, and released code for more details!.Project Page: Github: Thanks to all co-authors! @jiagengliu02 @zheyuanzhang99 @Siyuan_Zhou99 Reuben Tan @jw2yang4ai @du_yilun @gan_chuang .also thanks.

🎬 MindJourney in action. Given a spatial reasoning question.1️⃣ Imagine – VLM and world model “walk” the scene iteratively.2️⃣ Observe – the VLM picks up the clues from the tour.3️⃣ Answer – with context, the VLM replies. The imagination loop turns one frame into insight. 💡. 🧵5/

MindJourney achieves state-of-the-art results in multiple 3D spatial reasoning tasks. 🔥. Even OpenAI’s o1, which already has built-in inference-time reasoning capability, climbs higher once it plugs into MindJourney. The takeaway? Spatial imagination from a world model and a

Given a spatial reasoning query, MindJourney launches an iterative spatial beam search. 🔍. 1️⃣ The world model imagines a few quick ego-motions. 2️⃣ The VLM selects the branches that seem most useful for the question. 3️⃣ Use newly imagined viewpoints as additional evidence. Those

Recent VLMs can sketch images to “think,” but they’re stuck in toy grid-worlds. World models now simulate rich 3D physics and camera motion. Fuse one into a VLM and it gains a true imagination space—free to explore, perceive, and reason about real scenes. 🌍✨. 🧵2/

RT @du_yilun: VLMs often struggle with physical reasoning tasks such as spatial reasoning. Excited to share how we can use world models +….

Thanks @_akhaliq for sharing our work! . MindJourney fuses a world model with any VLM, so the model can first imagine walking around before it answers. From “one snapshot” to “what if I stand over there?”—and suddenly spatial reasoning hits SOTA. 🚀. Project Page:.

RT @_akhaliq: You can install anycoder as a Progressive Web App on your device. Visit and in the footer click set….

RT @ziqiao_ma: 📣 Excited to announce SpaVLE: #NeurIPS2025 Workshop on Space in Vision, Language, and Embodied AI! . 👉 .

I hope humans and robots live peacefully in the Virtual Community. Great work by @QinhongZhou !.#DetroitBecomeHuman #AI #Robotics

Nashville’s food is hands-down the highlight of CVPR for me so far. Sending a meat-lover’s salute to the South 🤤. P1 Hattie B.P2 Peg Leg Porker.#CVPR2025

Watched the notorious @celtics game while working on my NeurIPS submission. It took me 2½ hours to realized there’s something even more painful than rushing a NeurIPS paper. #Celtics #NeurIPS2025.

RT @_akhaliq: TesserAct is out on Hugging Face. Learning 4D Embodied World Models