Research questionHow do visual-space alignment and task balance determine whether joint VLM training helps or conflicts?A unified vision-language model must share representations across interpreting images and generating them. Compatibility between its visual input and output spaces, along with the balance of training data, can determine whether these capabilities reinforce or interfere with one another. Latest papersRecent research connected to this question, newest first.Cross-Task Generalization Between Understanding and Generation in Unified Vision-Language Models: A Controlled StudyThe evidence comes from controlled SmartWatch and modified CelebA benchmarks covering VQA, captioning, and text-to-image generation, using unified architectures built from SigLIP and VQ-VAE visual spaces. The study also varies visual-space distortions, task-data proportions, and attribute frequencies, and includes a real-case experiment with LLaVA. Its adapter analyses examine whether transfer arises from visual features or relationships learned by the base language model.research paper · Sep 4, 2026Uncovering Understanding-Generation Synergy in Native Unified Multimodal Models: From Representation, Task to SystemThe evidence concerns structurally native unified multimodal models trained without pretrained vision priors. It covers representation-, task-, and system-level interactions, including complex tasks requiring both capabilities and comparison with a matched planner–executor pipeline; conclusions are bounded to these settings.research paper · Sep 1, 2026