AI × Animals Weekly
Issue 1 · 13–19 November 2026
Subscribe
Two Commerson's dolphins surfacing in choppy water

Research & Data

New papers, open datasets and benchmarks, including studies that use research animals.

Researchers test whether image models can distinguish closely related species

PS-SPLIT tests category discovery on a task inspired by how taxonomists separate closely related species.

Note: Accepted at NeurIPS 2026 (Evaluations and Datasets Track), according to the authors’ arXiv listing. The benchmark uses existing bird photographs.

Lange and colleagues introduce PS-SPLIT, an image benchmark inspired by taxonomic splitting. Taxonomists sometimes find that populations previously grouped together actually represent distinct species. The benchmark uses this idea to construct tasks where a visual model must discover distinctions within a group.

The authors argue that widely used category-discovery benchmarks have become less useful for separating competing methods. Closely related animals can share many visual traits. In this setting a model must make sense of subtle variation, rather than relying on differences that are easy to see between broad categories.

The work provides an evaluation setup for comparing methods on the same problem. Visual clustering can suggest which images deserve closer inspection, while taxonomic decisions may also depend on genetic, geographic and other evidence.

Why it matters

Better image models could help biodiversity researchers identify animals that look alike. This benchmark tests that ability; it does not yet show a benefit for conservation in the field.

Get the weekly briefing

AI and animal news, in your inbox every Friday.