Don’t Read Everything: A Curvature-Conditioned Query for Linear Attention received the Best Paper Award at EMNLP 2026!