Complex KDA Enhances Linear RNN Expressivity via 2D Rotations
September 20, 2026
Complex KDA (CKDA) improves the expressivity of Kimi Delta Attention by combining a delta-rule transformation with a second reflection. By extending parameter ranges for gates and coefficients, the model can realize 2D rotations while maintaining the efficiency of linear RNNs.
HOW THIS AFFECTS YOU
●
researcherYou can improve the modeling capacity of linear recurrent architectures without increasing the computational complexity of updates.