Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery
NeurIPS 2026
Jowaria Khan · Anindya Sarkar · Yevgeniy Vorobeychik · Elizabeth Bondi-Kelly
Environmental discovery often requires learning from sparse observations while deciding where to collect new measurements under a limited acquisition budget.
OWL-GPS is a relevance-guided online learning framework that jointly performs sequential acquisition and online adaptation. It learns a region-specific relevance representation over environmental concepts and uses this shared representation both to decide which regions to query next and to guide online model updates.
The resulting loop is:
represent context → estimate relevance → acquire → adapt → recompute relevance → acquire again
We evaluate OWL-GPS primarily on real-world PFAS contamination discovery and additionally study its portability on a sparsified land-cover task.
See the project page and paper for full results.