Fig. 6

Few-shot classification performance on the Camelyon+²⁷. Benchmark. Boxplots show the performance of three models (Titan²¹, Prism²⁰, Chief¹⁸) under 2-way, 3-way, and 4-way classification settings, with varying numbers of shots (1, 2, 4, 8, 16, 32). (a–c) Accuracy, (d–f) Recall, and (g–i) F1-score across increasing few-shot levels. Each box represents model performance distribution over multiple test episodes. Titan and Chief generally outperform Prism in low-shot regimes, and the performance gap narrows with increasing shot numbers.