This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: Researchers at Cardiff University have presented two studies at the 2026 International Conference on Machine Learning (ICML 2026) that address fundamental challenges in modern AI: understanding the complex geometry and relationships within data and recognizing patterns that occur across very different scales. You Zhou, senior lecturer at Cardiff University's School of Medicine and senior author of the studies, said, "Biological research generates enormous amounts of complex data, but much of today's AI still struggles to understand the way biological systems are naturally organized.
Cells interact within tissues, and important patterns can appear at many different scales. "Our research addresses these challenges by developing AI methods that can better understand both the geometry of biological data and the relationships between local details and larger-scale structures." In the first paper, "VBA: Vector Bundle Attention for intrinsically geometric representation learning," the researchers introduce Vector Bundle Attention (VBA), a new AI architecture that rethinks how machines compare and interpret information. "Most current advanced AI systems are built on transformers, which use a mechanism to determine how different pieces of information relate to one another.
While highly effective, conventional attention mechanisms do not naturally account for the geometric structure that underlies many forms of complex data," added Zhou. The geometric limitations of AI are particularly important in biology, where individual cells are not isolated pieces of data. They exist within intricate molecular and spatial relationships that influence how tissues function and diseases develop.
The VBA model incorporates geometric relationships directly into the AI's attention mechanism, aligning information from cells according to their underlying geometry before making comparisons when modeling biological systems. To test the VBA model, the researchers used single-cell RNA sequencing and spatial transcriptomics—technologies that enable scientists to examine the molecular characteristics of individual cells and understand how cells are organized within tissues. The VBA model achieved state-of-the-art performance in single-cell RNA sequencing tasks, strong performance in spatial transcriptomics and competitive results on 3D datasets, highlighting its potential beyond biomedical applications.
In a second paper, "Dynamic Fractal Mamba: A neural renormalization group flow for scale-invariant sequence modeling," the second AI model, Dynamic Fractal Mamba (DF-Mamba), tackles another major challenge in AI—understanding information across dramatically different scales. "Many AI systems struggle to generalize when trained on relatively small data segments and then applied to much larger datasets. In biomedical research, this is a significant issue," said Zhou.
Inspired by concepts from physics used to describe how systems behave across scales, DF-Mamba repeatedly applies the same learned rules as it moves from smaller to increasingly larger patterns. This enables the model to integrate information within broader contexts efficiently. The DF-Mamba model learns from smaller-scale data and can successfully analyze much larger-scale datasets that it has never previously encountered, without requiring retraining.
Extract — continue reading at the source.
