SymbOmni: New AI model learns cumulatively, beats GPT-Image-1 by 40%

A team of researchers (Jinxiu Liu et al.) has unveiled SymbOmni, a new agentic omni-model designed to solve a fundamental flaw in current AI: the inability to learn cumulatively and evolve autonomously, which they call the 'perpetual novice' problem. Traditional monolithic models treat every task from scratch, wasting compute and failing to generalize. SymbOmni tackles this with a novel Symbolic Concept Box — an optimizable memory module that abstracts low-level operations into reusable Symbolic Workflow Instructions.

SymbOmni operates through an induction-transduction cycle: it first abstracts experiences into symbolic concepts (induction), then adaptively composes those concepts to solve new tasks (transduction). Training is done via verbalized backpropagation, using language-based feedback to enable continuous self-improvement without gradient-based fine-tuning. This allows the model to retain knowledge and evolve over time, much like a human expert.

In comprehensive experiments, SymbOmni significantly outperformed existing agent-based systems for iterative creation and even surpassed closed-source models like Nano Banana and GPT-Image-1 in both image quality and task success rates. Additionally, it reduced token consumption by over 40% while maintaining competitive generation quality. On continual learning benchmarks, SymbOmni set a new state of the art, demonstrating cumulative gains — a key leap toward lifelong learning AI.