Scientists often struggle to identify rare anomalies in images because standard AI models are trained on vast amounts of common examples, leaving them blind to the few critical cases they need most. While foundation models have revolutionized image recognition for everyday photos, researchers found that simply fine-tuning these powerful tools on scientific data yields limited gains when dealing with this imbalance. The breakthrough comes from realizing that penultimate-layer features hold the key to spotting those rare occurrences. By combining deep insights from both penultimate- and final-layer stages of the neural network, a new framework called SciLT successfully balances performance across common and rare classes. This approach establishes a robust baseline for adapting AI to scientific domains where data distributions shift dramatically, offering a practical solution for detecting elusive phenomena in research. Source: SciLT: Long-Tailed Classification in Scientific Image Domains by Jiahao Chen, Bing Su, https://arxiv.org/abs/2604.03687