publications

2026

1. ECCV

   SymbOmni: Evolving Agentic Omni Models via Symbolic Concept Learning

   Jianru Li^*, Jinxiu Liu^*, Tanqing Kuang^*, and 4 more authors

   2026

   Abs Bib

   We propose SymbOmni, which solves the continuous learning bottleneck in visual generation by introducing a Symbolic Concept Box and Verbalized Backpropagation. It achieves state-of-the-art on AIGC benchmarks with over 40% token reduction, outperforming mainstream closed-source models. The paper was accepted by ECCV 2026.

   @article{li2025symbomni,
     title = {{SymbOmni}: Evolving Agentic Omni Models via Symbolic Concept Learning},
     author = {Li, Jianru and Liu, Jinxiu and Kuang, Tanqing and Liu, Xuanming and Mei, Kangfu and Wen, Yandong and Liu, Weiyang},
     year = {2026},
   }