Validation-Aligned Coreset Selection for Budgeted Few-Shot Classification

<title>Abstract</title> <p>Few-shot classifiers can be dominated by the initial choice of which labelled examples are retained for training. We study this choice as a training-only algorithm-selection problem. Validation-Aligned Coreset Selection (VACS) evaluates a finite portfolio of class-balanced selectors on internal validation splits and then rebuilds the final coreset with the selected rule on the full training split. The portfolio contains random selection, herding, k-center, boundary selection, k-means medoids, and a margin-weighted representative selector. Across five public benchmarks, 80 matched low-budget dataset-seed-budget units, and a distance-weighted 3-nearest-neighbour classifier, the fast single-split protocol VACS-F reaches 70.6% mean accuracy, numerically tying the strongest pooled static selector. The repeated protocol VACS-R averages five internal validation rankings and reaches 72.1%, improving over herding by 1.54 percentage points with a 95% bootstrap interval of 0.68 to 2.49 points. On a larger Covertype audit, VACS-R improves over herding but exactly matches the best static rule, MARC. Boundary-only, BADGE-inspired, frozen text-embedding, and frozen image-embedding checks define the method's limits. VACS supports a narrow conclusion: repeated training-only validation can make selector choice competitive under extreme labelled budgets, but it is not a broadly superior coreset rule.</p>