AI Research Advances: Agentic Systems, Multimodal, and Reasoning

Multiple strong research papers released: AutoScientists (decentralized agent teams for scientific experiments, +8.33% BioML-Bench, 1.9x faster GPT training); OSP-Next (efficient video generation with sparse attention, 83.73% VBench, 1.64x speedup); NEO-ov (native one-vision model, end-to-end pixel-word learning, narrows gap to modular VLMs); ResearchMath-14K (scaling research-level math reasoning, reveals hallucination risk in newer models); Gamma-World (multi-agent world models from NVIDIA, 24 FPS real-time). These signal rapid progress in agentic AI, multimodal understanding, and reasoning capabilities.