The recently proposed learning-to-learn iterative search algorithm (LISA) for multiple-input multiple-output (MIMO) detection can achieve excellent BER performance for low-ordermodulation signals, such as BPSK/QPSK, but its performance is degraded as the modulation orderincreases. This limits its application in high data rate transmission. Motivated by this observation,this paper proposes a decomposition-ordering-based LISA (DO-LISA) to improve the accuracy ofLISA for higher-order QAM signals. Firstly, the detection of high-order QAM signal is transformedinto the detection of QPSK signal to exploit the facility of binary classification. To further exploitthe sequential features of a bidirectional long short-term memory (BiLSTM) network, channelordering is adopted to enhance the detection accuracy. Moreover, to make the proposed methodmore practical, low-rank decomposition and channel pruning are adopted to accelerate and compressthe neural network. Numerical experiments show that the proposed method can achieve quasi-MLBER performance. Compared with the full model, the compact model achieves 2.37× compressionratio and 2.38× acceleration ratio with limited accuracy loss. In addition, the proposed DO-LISA shows its good generalization ability by preserving very good BER performance both inthe correlated channel condition and imperfect channel condition.