SmartApe AI

Oct 19, 2026 · Artificial Intelligence

Visual Generation Knowledge Fails to Transfer to Text QA: New Research Bridges Gap

A new paper from Tsinghua and Zhipu AI introduces Knowledge Transfer Probing (KTP) to test if unified multimodal models can use knowledge learned from image generation for text QA, revealing a significant gap; interleaved image-text alignment training and visual chain-of-thought reasoning substantially improve text QA accuracy across BAGEL, SenseNova-U1, and GLM-Omni, though a gap persists.