"Test-Time Spectral Adaptation for Image-to-Image Multispectral Retrieval," by Tien-Anh Nguyen, Thanh-Hai Tran and Dr. Xuan-Bach Le, has been accepted at ACIVS 2026 in Okinawa, Japan, December 2026. Sentinel-2 imagery carries visible, red-edge, near-infrared and short-wave infrared bands, while models such as CLIP accept only three channels. The authors propose a test-time adaptation scheme that leaves every backbone parameter frozen and updates just B band-weight logits per image. Those logits start uniform, are refined by a lightweight manifold-consistency objective guided by a text prior built from class names, and yield attribution scores at the level of individual bands. On 5-fold EuroSAT retrieval with disjoint query and gallery splits, Recall@1 rises from 80.28% for RGB-CLIP to 86.49%, with mAP at 46.50%. On BigEarthNet-MM across five seeds, the method matches full-band averaging within one standard deviation, but adds per-band interpretability and near-zero adaptation overhead. Paper: https://lexuanbach.github.io/publication/ACIVS2026c.pdf Supplement: https://lexuanbach.github.io/publication/ACIVS2026c-supplement.pdf Code: https://github.com/tienesng06/acivs_supplementary