LOCUS: Local Visual Cue Search for Enhancing Fine-Grained Perception in Multimodal Large Language Models
Researchers have introduced two new benchmarks and training frameworks to address limitations in multimodal large language models (MLLMs). GePBench focuses on evaluating and improving MLLMs' fundamental geometric perception abilities, revealing significant deficiencies in current state-of-the-art models. Separately, the LOCUS framework enhances fine-grained visual perception by training MLLMs to better utilize local visual cues within an image, combating "visual context rot." AI
IMPACT These advancements aim to improve the reliability and capabilities of multimodal AI systems in understanding complex visual information.