Introducing the Describe Anything Model (DAM), a powerful Multimodal LLM that generates detailed descriptions for user-specified regions in images or videos using points, boxes, scribbles, or masks. Open-source code, models, demo, data, and benchmark at:

RT @LongTonyLian: Excited to share that Describe Anything has been accepted at ICCV 2025! 🎉. Describe Anything Model (DAM) is a powerful Mu….

RT @hanna_mao: We build Cosmos-Predict2 as a world foundation model for Physical AI builders — fully open and adaptable. Post-train it for….

RT @googleaidevs: Introducing Gemini CLI, a light and powerful open-source AI agent that brings Gemini directly into your terminal. >_. Wri….

RT @ComfyUI: 🎉 ComfyUI now natively supports NVIDIA’s Cosmos-Predict2 model family!. Cosmos-Predict2 brings high-fidelity, physics-aware Im….

RT @hanna_mao: 🚀 We're releasing Cosmos-Predict2 — our developer-first, top-performing world foundation models for Physical AI!.🔗 https://t….

RT @ashawkey3: Happy to share our work PartPacker:.We enable one-shot image-to-3D generation with any number of parts!. Project page: https….

RT @chenhsuanlin: Cosmos-Predict2 is our latest open video foundation model for Physical AI!. If you’re at #cvpr202….

RT @TsungYiLinCV: The physics meets vision workshop just started! Come joining us!

RT @prithvijitch: The WorldModelBench workshop is happening tomorrow (June 12th) at #CVPR2025! We have an exciting series of talks, do atte….

If you are attending #CVPR2025 tomorrow, please visit two highly relevant workshops organized by our team members:.- Vision Meets Physics: - Benchmarking World Models:

RT @qsh_zh: 🚀 Introducing Cosmos-Predict2!. Our most powerful open video foundation model for Physical AI. Cosmos-Predict2 significantly im….

RT @FangyinWei: Join us on the 1st workshop on Vision Meets Physics: Synergizing Physical Simulation and Computer Vision at #CVPR2025 tomor….

RT @mli0603: Cosmos-Reason1 has exciting updates 💡.Now it understands physical reality — judging videos as real or fake! Check out the reso….

This work is led by our amazing intern @chenhuay17 and his host @Haoxiang__Wang!. Other collaborators: @zkwthu @qsh_zh @charlesfornlp @haotian_yeee @TsungYiLinCV @liu_mingyu Jun Zhu. The code and model will be released soon!.

As a supervised learning method, NFT outperforms leading RL algorithms like GRPO and DAPO in 7B model experiments and performs similarly to DAPO in 32B settings. We conduct 3-4 independent experiments in 7B training to reduce random noise.

Why is it even possible to optimize your LLMs on negative data via supervision? The key is a tight coupling between positive & negative answer distributions generated by the same LLM in online training.

We proved that NFT is equivalent to GRPO in strict on-policy training, despite their entirely different theoretical foundations. The normalized advantage design of GRPO is implicitly reflected in NFT's objective. NFT bridges the existing RL and SL theoretical framework.

Traditional supervised learning methods like rejection sampling fine-tuning cannot leverage negative feedback. The key to NFT’s success lies in its effective use of negative feedback, by constructing an implicit negative policy parameterized by the target LLM. Training this

Is self-improvement exclusive to RL?. Can we use supervised learning to match LLMs trained with SOTA RL algorithms?. In Negative-aware Fine-Tuning (NFT), we introduce a purely supervised learning method to enhance LLMs' math reasoning with no external teachers. NFT matches or

RT @OriolVinyalsML: Today we introduced Gemini Diffusion⚡️ (& DeepThink, Veo3, Imagen4, 2.5 updates. ). It's been a dream of mine to remo….