Cross-Modal Retrieval

Haiku achieved strong global patch-level cross-modal retrieval on held-out paired data. Zero-shot 1-nearest-neighbor patch annotation favored Haiku over MUSK and random baselines. Text-to-mIF retrieval was weaker than image-to-image retrieval but still above baseline. Retrieval evaluation used global held-out galleries across all 336 paired test slices rather than within-slice restricted retrieval. The article interprets strong H&E-to-mIF and mIF-to-H&E retrieval as evidence of learnable correspondences between morphology and spatial protein organization.