NeurIPS 2026｜DGP integrates ViT and LLM layers, achieving a new breakthrough in VLM architecture

DGP achieves full integration between ViT and LLM layers, enabling large models to process image information as needed at each layer. In the NeurIPS 2026 paper, DGP correctly read two prices and calculated 14.80 euros in the Funghi pizza total price task, while LLaVA-1.5 misread the prices and generated hallucinations. This architecture supports 175B parameter scale with a context length of 128k tokens and provides closed-source services through an API.