PhDone :) – at Ann Arbor, MI

NEPA: Next-Embedding Predictive Autoregression A simple objective for visual SSL and generative pretraining. Instead of reconstructing pixels or predicting discrete tokens, we train an autoregressive model to predict the next embedding given all previous embeddings. Key ideas: - One self-supervised signal: cosine-style next-embedding prediction - Autoregression runs directly on the embeddings from a native encoder (no offline encoder) - No pixel decoder (and loss), no contrastive pairs, no task-specific heads, no random masks Scales into modern ViT backbones and stays competitive after supervised fine-tuning: - ImageNet-1K (Base 83.8%; Large 85.3%) - ADE20K Fully open-sourced with reproducibility verified: - Homepage: sihanxu.me/nepa/ - Paper: arxiv.org/abs/2512.16922 - Code: github.com/SihanXU/nepa - Weights: huggingface.co/collections/S… This work is led by @6SihanXu and advised by @SLED_AI, @sainingxie, and Stella X. Yu. Contributors: me, @wenhaocha1, @ChenXuweiyi, and @JinWeiyang18434.

Will be at #NeurIPS2025 (San Diego) Dec 1-9, then in the Bay Area until the 14th. Hmu if you wanna grab coffee and talk about totally random stuff. Thread with a few things I’m excited about. P.S. 4 NeurIPS papers all started pre-May 2024 and took ~1 year of polishing...so proud of the team!

Still wrapping up a few reality-check experiments and polishing the tutorial structure ... but we're excited! P.S. Sadly the ARC-AGI team can't join the tutorial panel this time due to conflict of schedule, but they’ll be with us at the @LAW2025_NeurIPS later in the NeurIPS program. Stay tuned :)

Over the past few months, I’ve heard the same complaint from nearly every collaborator working on computational cogsci + behavioral and mechanistic interpretability: “Open-source VLMs are a pain to run, let alone analyze.” We finally decided to do something about it (thanks @fredahshi @jzhou_jz for initiating and co-leading this). VLM-Lens (to appear in #EMNLP2025 demo) is a toolkit for systematically analyzing and interpreting open-source VLMs. [Code] github.com/compling-wat/vlm-… [Paper] arxiv.org/abs/2510.02292 VLM-Lens abstracts away model-specific complexity and gives you fine-grained access to any internal representation across 16 VLMs and their 30+ variants. It is built on these simple principles: 1. Unified access: consistent API across all models 2. YAML-configurable: minimal model-specific code needed 3. Extensible: add new models in a few lines 4. Interpretability-ready: plug into probes, PCA (more under dev) Let’s make VLM research a little easier to see through. :)

Excited to announce the #NeurIPS2025 Workshop on Bridging Language, Agent, and World Models for Reasoning and Planning (LAW) sites.google.com/view/law-20… The LAW 2025 workshop brings together Language models, Agent models, and World models (L-A-W). It aims to spark bold conversations around whether LLMs possess internal world models, how we can build more grounded and generalizable agents, and what it takes to go beyond today’s capabilities. Huge thanks to our amazing co-organizers and speakers, and to @LambdaAPI for generously sponsoring us! Join us at LAW 2025 as we chart the future of AI systems that can reason, act, and adapt in complex, ever-changing environments.

Unfortunately, I’ll be missing #ACL2025NLP this year — but here are a few things I’m excited about! 👇 Feel free to DM me if you’d like to chat.

📣 Excited to announce SpaVLE: #NeurIPS2025 Workshop on Space in Vision, Language, and Embodied AI! 👉 space-in-vision-language-emb… 🦾Co-organized with an incredible team → @fredahshi · @maojiayuan · @DJiafei · @ManlingLi_ · David Hsu · @Kordjamshidi 🌌 Why Space & SpaVLE? We never directly “see” space. Instead, we reconstruct it, describe it through language, and navigate its constraints to plan our actions. Let’s bring together communities to tackle spatial intelligence across 2D/3D reasoning, grounded language, and real-world robotic planning. 📝 Call for Papers • 4-page shorts or 9-page fulls (non-archival) • Topics: spatial representation, grounding, datasets/benchmarks, foundation models & more. 🗓️ Key Dates • Deadline: Aug 22 • Notifications: Sep 22 • Camera-ready: Oct 25 Mark your calendars & start drafting! 🚀 🎤 Star-studded keynotes spanning CogSci, NLP, CV, Robotics. Amir Zadeh · Barbara Landau · Dieter Fox · Joshua Tenenbaum @MITCoCoSci · Joyce Chai @SLED_AI · Ranjay Krishna @ranjaykrishna · Saining Xie @sainingxie. Can’t wait for the insights! 🏆 Best Paper = $3 k cloud credits + Runner-up $1.5 k. Huge thanks to our sponsors: Lambda (@LambdaAPI), Alquist Robotics (@alquistrobotics), and EdenSign (@Edensign_ai). More sponsors welcome! DM us if you’d like to support spatial-AI research. Join us in San Diego to push the frontiers of spatial understanding and reasoning across CV, NLP, and robotics!

Thrilled to share that VEGGIE is accepted to #ICCV2025! 🎉 Check out the full thread by @shoubin621 for details. Funny enough — it’s been 6 years since I came to the US, and this might be my first time setting foot in Hawaii. 🌴

Our study on pragmatic generation is accepted to #COLM2025! Missed the first COLM last year (no suitable ongoing project at the time😅). Heard it’s a great place to connect with LM folks, excited to join for round two finally.

Can we scale 4D pretraining to learn general space-time representations that reconstruct an object from a few views at any time to any view at any other time? Introducing 4D-LRM: a Large Space-Time Reconstruction Model that ... 🔹 Predicts 4D Gaussian primitives directly from multi-view tokens (no motion vectors, no HexPlane); 🔹 Uses a clean, minimal Transformer backbone; 🔹 Generalizes with fast, high-quality feedforward rendering at any view and infinite frame rate. Check out more interactive demos and scaling behaviors on our homepage/paper. 👉Website: 4dlrm.github.io/ 👉Paper: arxiv.org/abs/2506.18890