#computer vision
Nov 2025
Does Understanding Inform Generation in Unified Multimodal Models? From Analysis to Path Forward
UniSandbox, a decoupled evaluation framework paired with controlled, synthetic datasets to avoid data leakage and enable detailed analysis, reveals a significant understanding-generation gap, which is mainly reflected in two key dimensions: reasoning generation and knowledge transfer.
Yu-Wei Niu, Weiyang Jin, Jiaqi Liao et al.
· arXiv.org · 10 citations