Evaluating xLSTM for structural coherence and efficient recurrent inference

Long-form symbolic music generation requires models to maintain both local musical plausibility and large-scale structure over long sequences. While Transformer-based systems such as Museformer have shown strong long-context modeling ability, they also incur substantial decoding cost. In this work, we investigate the Extended Long Short-Term Memory (xLSTM) architecture as an alternative for long-form symbolic music generation. We train xLSTM on the Lakh MIDI Dataset using REMIGEN encoding, implement recurrent-state inference for efficient autoregressive decoding, and evaluate the model both comparatively and through xLSTM-specific long-sequence analysis. Our comparative framework combines musical quality metrics, human listening tests, and three self-similarity-matrix-based structural coherence metrics: Block Coherence, Repetition Density, and Off-diagonal Similarity. We also introduce Decode Memory Growth Rate (DMGR), a rate-based metric for comparing memory growth during autoregressive decoding across architectures. The results show that xLSTM is the closest model to human music on all three structural coherence metrics, achieves the highest Mean Opinion Score, wins pairwise preference tests against Museformer and Lookback RNN, and is most often mistaken for human-composed music in the Turing test. On the efficiency side, recurrent-state decoding yields up to 25× speedup over the original parallel-style generation loop, and xLSTM maintains near-constant decoding memory growth of 0.49 MB per 1k generated tokens, compared with substantially higher memory growth for Museformer. Additional analysis shows that xLSTM extrapolates beyond its 4,096-token training context, although grammar errors rise gradually at longer lengths. Overall, the study shows that xLSTM is a viable architecture for long-form symbolic music generation and offers a strong practical trade-off between structural quality and decoding efficiency.