LATTE: Learning to Think with Vision Specialists (2412.05479v3)

Abstract: While open-source vision-LLMs perform well on simple question-answering, they still struggle with complex questions that require both perceptual and reasoning capabilities. We propose LATTE, a family of vision-LLMs that have LeArned to Think wiTh vision spEcialists. By offloading perception to state-of-the-art vision models, our approach enables vision-LLMs to focus solely on reasoning over high-quality perceptual information. To train LATTE, we synthesize and filter a large dataset of 273K multi-modal reasoning traces over perceptual outputs of vision specialists. LATTE trained on this data achieves significant gains over baselines across 6 benchmarks covering both perception and reasoning abilities. Ablation studies reveal that the effectiveness of multi-modal reasoning traces depends on the data sources, formats, and quality of thoughts.