In "A Categorical Atlas of Parameter-Efficient Fine-Tuning: Cones, Orbits, and Gauges" (arXiv:2610.12728), the authors argue that the usual way of comparing parameter-efficient fine-tuning methods is too narrow. Hundreds of methods are typically ranked by parameter count and benchmark score, but the preprint suggests a different lens: compare them as maps.
The setup starts with a pretrained checkpoint as a base point in weight space. A weight-space method is then understood as something acting from that point, rather than as an isolated set of tuned parameters. The title points to categorical tools—cones, orbits, and gauges—as a way to organize the space of methods, although the released abstract does not yet define these terms in detail.
Because this is a single preprint and the abstract is brief, the article does not attempt to summarize experimental results or compare the framework with existing taxonomies. Its significance is conceptual: it offers a geometric and categorical alternative to benchmark-driven comparisons for a crowded field.