Brief

Researchers have developed a novel method called SPORT (Step-wise Preference Tuning) to train multimodal agents without relying on extensive human-annotated data. This approach uses an iterative process of task synthesis, step sampling, step verification, and preference tuning to enable agents to autonomously discover effective tool usage strategies. Evaluations on the GTA and GAIA benchmarks demonstrated significant improvements in agent performance, highlighting the method's generalization capabilities. AI

Researchers have developed a new framework using a high-fidelity game engine to generate perfectly paired driving datasets. This method modifies illumination and weather conditions while keeping scene geometry and object placement consistent. The generated datasets allow for a more accurate analysis of how photometric shifts affect visual perception systems in autonomous driving, as demonstrated by an evaluation of semantic segmentation models. AI

Researchers have introduced GTA, a new method for generating 3D worlds from single images. Unlike previous approaches that often prioritize appearance over structure, GTA first generates the geometric layout of a scene and then synthesizes its appearance. This two-stage video diffusion model process aims to improve structural fidelity and cross-view consistency. Experiments show GTA outperforms existing methods in accuracy and visual quality, and can also enhance other 3D generation pipelines. AI