AI News — Tuesday, March 24, 2026
OpenAI details its ongoing efforts and strategies to ensure the safe and responsible deployment and use of its text-to-video model, Sora.
Researchers introduce HopChain, a novel method for synthesizing multi-hop data to improve the generalizability of vision-language reasoning models.
Air Street Capital, a venture capital firm focused on AI, has successfully raised a $232 million fund, positioning it as one of Europe's largest solo VCs.
This paper presents Astrolabe, a technique that leverages forward-process reinforcement learning to enhance the distillation of autoregressive video models.
A novel approach using λ-Calculus is presented to address the 'long-context rot' problem in Large Language Models, improving their ability to handle extended contexts.
A unique benchmark proposes evaluating LLMs by having them write code to control units in a 1v1 real-time strategy game.
An introductory guide explains the critical importance of prompt engineering and how the way questions are phrased significantly impacts AI model responses.
TerraScope introduces a new framework for pixel-grounded visual reasoning specifically designed for complex Earth observation tasks.
This paper proposes ProactiveBench, a new benchmark to evaluate and improve the proactive capabilities of multimodal large language models.
Researchers introduce FlowScene, a method for generating style-consistent indoor scenes using multimodal graph rectified flow.
LumosX is a new system designed to generate personalized videos by effectively relating various identities with their specific attributes.
A new subgoal-driven framework is proposed to significantly improve the performance and planning capabilities of LLM agents in long-horizon tasks.
A novel method is introduced that allows for versatile editing of video content, actions, and dynamics without requiring additional training.
This research explores the concept of 'Hyperagents,' advanced AI agents capable of complex, adaptive behaviors in dynamic environments.
BEAVER offers a training-free hierarchical prompt compression method that uses structure-aware page selection to efficiently manage prompts.