#visual reasoning

ByteDance introduces VGR, a multimodal large language model that dynamically integrates visual information during reasoning, significantly improving accuracy and efficiency in vision-language benchmarks.

VeBrain introduces a unified multimodal AI framework that combines visual reasoning and robotic control, achieving superior performance in complex real-world robotic tasks.

The PHYX benchmark uncovers key weaknesses in current multimodal AI models’ ability to perform physical reasoning, emphasizing the challenge of integrating visual data with symbolic and causal knowledge.