ARIMA: Reconstruction-Grounded Predictive Representation Learning for Symbolic Music

2026-07-10

AI Summary

Existing self-supervised representation methods for symbolic music are limited by tokenizer dependency and the absence of explicit window-level temporal modeling, making it challenging to balance local detail preservation with global dynamics. This work proposes ARIMA, a framework that encodes fixed-duration segments into continuous latent representations and employs causal contrastive learning to predict the next temporal window. By incorporating a structured reconstruction task to regularize the encoder, ARIMA learns compact, temporally coherent, and detail-preserving representations without requiring explicit variance regularization. The method significantly outperforms current baselines on tasks such as harmonic analysis, rhythmic modeling, and cross-performance retrieval, and matches or exceeds the performance of substantially larger models on various other music understanding benchmarks, demonstrating its efficiency and strong generalization capability.

Abstract

Self-supervised learning for symbolic music has advanced largely through token-level pretraining, but such representations remain tied to tokenizer-specific sequences and often provide time-span-level embeddings only indirectly. In this paper, we propose ARIMA, a reconstruction-grounded latent predictive framework for symbolic music that learns compact window-based representations directly from data. ARIMA encodes each fixed-duration window into a continuous latent representation, trains a causal predictor with contrastive next-latent prediction, and grounds the encoder through structured reconstruction of music elements. This design preserves local musical details while modeling temporal progression across windows. We evaluate ARIMA on downstream tasks spanning various levels of music understanding. Results show that ARIMA is particularly efficient and effective on tasks involving harmonic, timing, and cross-performance retrieval, while remaining competitive with much larger baselines on other tasks. Ablations further show that next-latent prediction is essential for temporally integrated representations, and that structured reconstruction stabilizes latent learning without requiring explicit variance regularization. The code is at https://github.com/AndyWeasley2004/symbolic_music_wm.

Problem

• symbolic music
• self-supervised learning
• predictive representation
• window-based representation
• token-level pretraining

Innovation

• self-supervised learning
• symbolic music
• latent predictive modeling
• structured reconstruction
• contrastive next-latent prediction