🚀 Training an image generation model and picking sides between autoregressive (AR) and diffusion? Why not both? Check out MADFormer with half of the model layers for AR and half for diffusion. AR gives a fast guess for the next patch prediction while diffusion helps refine the

RT @yuqirose: How do we ground #LLMs for Scientific Problems to mitigate the issue of hallucination? Check out our #icml2025 paper on ``Ad….

🎯 Takeaway: Tune your AR⇄Diffusion balance to match your compute budget, more AR for speed and more diffusion for quality. 🙏 Huge thanks to @XiaochuangHan and Yulia @tsvetshop for the collaboration! Keen to hear your thoughts!. ( 4 /🧵).

💡 Key findings:. - Under a low inference budget/NFE, simply allocating more layers to AR and less to diffusion (i.e., an AR-heavy design) yields up to 75% FID gains; with high inference budget, diffusion-heavy designs win for finer details. - More on optimal AR block size,

@tsvetshop @XiaochuangHan 🔍 MADFormer explores the Mixed AR + Diffusion design space in image generation Transformers. - Token axis: split an image into blocks of patches (e.g. 4/16/64 blocks), with AR conditioning across the blocks and diffusion within the block. - Layer axis: allocate the first (N–D)

RT @stingning: Our framework supports various online RL algorithms. In our experiments, we use GRPO with the following optimizations:.1️⃣ P….

RT @stingning: SOTA: SimpleVLA-RL achieves 98.4% on LIBERO. 🎯With only 1 trajectory/task for SFT:. 🚀LIBERO-Avg: 48.9%→94.1%. 🚀LIBERO-Long….

RT @stingning: Moonlighting a bit: we implement Online RL for VLA models with @verl🤖, and find simple outcome rewards can work surprisingly….

RT @ZhiyuanZeng_: 🛠️ Build your own LLM (or benchmark) analysis/debugging tool 🚀 Try our demo .

RT @ZhiyuanZeng_: Is a single accuracy number all we can get from model evals?🤔.🚨Does NOT tell where the model fails.🚨Does NOT tell how to….