Document is current

Article History

Received: 30 May 2025

Accepted: 28 January 2026

First Online: 7 April 2026

Declarations

:

: We assume that a correct and known hierarchy is available. While it is possible to use LLM-generated hierarchies (Liu et al., ), verifying the correctness and usability is an exciting direction for future work. In this work, we focus on hierarchy-aware OOD generalization, i.e., structured generalization within known taxonomies, while broader domain-level generalization is beyond the current scope. The formulation of Balance Hyperbolic Learning is very well suited for discriminative tasks such as OOD detection. However, this design may limit the capacity to capture fine-grained semantic relations among nodes within the same level, particularly in tasks where intra-level similarities are context-dependent. Learning hierarchies jointly with the task, or dynamically refining them from data, remains an exciting research direction.