In a groundbreaking development, researchers from MIT have introduced a novel training method that significantly enhances vision-language models’ ability to locate personalized objects in new scenes. This advancement addresses a critical limitation in generative AI, where models like GPT-5 excel at recognizing general objects but struggle with identifying specific items, such as a pet among many others. By leveraging carefully curated video-tracking data, the new approach allows these models to learn from context rather than relying solely on pre-existing knowledge.

The advent of generative AI has revolutionized how we interact with data, but a significant gap remains in predictive analytics. Kumo’s Relational Foundation Model (RFM) aims to bridge this divide by applying the zero-shot capabilities of large language models (LLMs) to structured databases. This innovative approach allows businesses to predict outcomes like customer churn or fraud detection without the traditional bottlenecks of manual feature engineering. Kumo’s RFM transforms relational databases into interconnected graphs, enabling the model to learn complex relationships across multiple tables seamlessly.