SymbOmni

Agentic omni-model research on symbolic concept learning and continual visual generation.

SymbOmni explores how agentic omni-models can evolve through symbolic concept learning. The project focuses on breaking the continual learning bottleneck in visual generation.

Key points:

  • Co-first author project with Jianru Li, Jinxiu Liu, and Tanqing Kuang
  • Introduces a symbolic concept box and verbalized backpropagation
  • Achieves strong AIGC benchmark performance with more than 40% token reduction
  • Revised manuscript is currently under review