Vision-language models (VLMs) have become strong zero-shot classifiers, but most benchmarks focus on broad categories. A new paper on arXiv argues that fine-grained tasks and large hierarchical label spaces remain understudied, and asks whether a more structured approach—reasoning through a tree of labels—can help.
The authors investigate decomposing visual classification into tree-based reasoning steps, where a model navigates a hierarchy of possible classes rather than predicting from a flat list. This mirrors how humans often narrow down options, and could be especially useful when the number of labels is very large or the differences between classes are subtle.
The paper is an assessment of this idea, probing whether VLMs benefit from explicit tree structure during classification. As the abstract does not include final results, the key contribution at this stage is the framing and evaluation of tree-based reasoning as a method for improving fine-grained visual recognition.