MetaX Weekly AI Paper Review Week 26 2025 -- Innovative Approaches to Zero-Shot Parameter Generation Image Generation Document Understanding and Ultra-Long Generation, AI Technology Development Across Diverse Domains Including Interactive Game Worlds and 3D Animation: Drag-and-Drop LLMs: Instead of fine-tuning for each individual task, directly generates model weights from prompts alone, dramatically reducing training costs. Light of Normals: Through integrated representation separating lighting and surface shape features, accurately restores 3D surface of objects in arbitrary lighting environments. Vision-Guided Chunking Is All You Need: Improves accuracy of RAG systems by understanding and splitting document visual structure using large-scale multimodal models. OmniGen2: A versatile generation model with separated decoding paths for text and images, processing diverse tasks including text-to-image conversion and editing with a single model. ShareGPT-4o-Image: Large-scale dataset mimicking GPT-4o image generation capability for training open-source multimodal models. Additional papers covered: interactive game world generation enabling AI to create playable environments from descriptions; 3D animation synthesis from single-view video; and advances in long-document generation maintaining coherence over very long outputs.