PHOEBI: An Open-World Benchmark for Bacterial Identification in Phase-Contrast Microscopy
The PHOEBI dataset tests bacterial identification using optical microscopy across 40 species combinations. The study found that three lightweight decoders—SIMPLEXUNMIX, PROTOMATCH, and CHANNELGROUP—successfully avoid a significant performance drop ('compositional collapse') when tested on unseen species combinations. This means these methods can robustly identify known bacteria and discover entirely new ones without needing further training.