Meta-Learning for Domain Generalization Under Distribution Shift: Methods, Benchmarks, and Open Challenges

Domain generalization (DG) explores how a model trained on a fixed set of source domains can perform reliably on unseen tar-get domains. Meta-learning addresses this by training models to explic-itly learn to generalize, yet curated benchmarks have repeatedly shown that well-tuned evidence-based risk minimization (ERM) is a formidable baseline. This survey examines that tension with a focus on what ac-tually generalizes and why. We contribute three things. First, a bivari-ate taxonomy that cross references shift type (covariate, conditional, in-variant covariate, category, compound) against meta intervention level (data/augmentation, representation/gradient, optimization/parameter, prompt/foundation). Second, a structured comparison of benchmark fam-ilies: DomainBed, WILDS, single source, and open set settings that shows how benchmark choice, and not just method design, drives published conclusions. Third, a critical analysis of conditions under which meta-learning gains over ERM are real versus aphemeral, updated for the 2023-2026 period when foundation models, causal approaches, and meta prompting have substantially changed the landscape. We conclude with actionable open challenges and directions where meta-learning retains a genuine edge.