AI Highlights from March 2025

• Gemini Robotics introduces advanced vision-language-action (VLA) models that enable robots to perform complex tasks with humanlike reasoning and physical interaction, significantly enhancing their utility in real-world applications.

• Gemma 3 has outperformed Deepseek v3 in the Arena, achieving this feat with only 1 GPU compared to Deepseek's 32 GPUs.

• Block diffusion language models bridge the gap between autoregressive and diffusion models, enhancing flexible-length generation and inference efficiency through techniques like KV caching and parallel token sampling.

• GS-Cache is an innovative framework that enhances 3D Gaussian Splatting (3DGS) by integrating a cache-centric pipeline and optimized rendering techniques, achieving a 5.35x performance improvement and supporting 2K rendering at over 120 FPS.

• The SEA-VL dataset addresses the need for culturally representative vision-language data in Southeast Asia, utilizing crowdsourcing, web crawling, and AI image generation to achieve this goal, with web crawling yielding ~85% cultural relevance at a lower cost.

• Cost-optimal LLMs can be achieved by optimizing context length and attention head configuration, revealing that larger models with fewer heads yield lower computational costs and better performance on long sequences.

• Gemini 2.0 Flash now enables developers to generate images through multimodal input, enhancing storytelling and image editing capabilities, available for experimentation globally via the Gemini API.

• Gemma 3 is a lightweight, state-of-the-art open model that excels on single GPU or TPU setups, outperforming competitors like Llama-405B in preliminary evaluations.

• NVIDIA's latest neural rendering tools and DLSS 4 are set to revolutionize game development, enabling developers to create photorealistic graphics and immersive experiences with unprecedented efficiency.

• GTC 2025 will showcase cutting-edge advancements in AI, robotics, and accelerated computing, featuring influential speakers like Yann LeCun and Frances Arnold, who will challenge conventional thinking and inspire innovation.

• This study introduces SR$^3, a novel approach that utilizes LLMs' hallucination to generate negative reasoning, enhancing fake news detection by learning from both correct and incorrect interpretations of news content.