symbolic sequence representation

This study investigates whether large language models (LLMs), trained exclusively on text without explicit musical supervision, can implicitly acquire representations of symbolic music structure. Methodologically, we prompt LLMs with plain-text instructions to generate MIDI sequences, construct a symbolic music dataset from these outputs, and train dedicated neural networks for downstream tasks—including music genre classification, style identification, and melody completion. We provide the first empirical evidence that LLMs develop internal representations of core musical concepts—such as pitch, rhythm, and tonality—as demonstrated by the viability of their generated MIDI data for supervised learning: classification accuracy approaches that of fully supervised baselines, and melody completions exhibit musically coherent phrasing and harmonic consistency. Results indicate that LLMs possess nascent—but incomplete—reasoning capabilities over symbolic music structure, establishing a novel paradigm for zero-shot and few-shot music generation and understanding.

This paper proposes a label-free, unified modeling paradigm for symbolic music, framing diverse tasks—including chord recognition, melody generation, and drum pattern generation—as a single “music-to-music” sequence modeling problem. Methodologically, it introduces (1) a novel *function alignment* paradigm that implicitly encodes functional semantics (e.g., harmonic role, rhythmic function) via intrinsic structural patterns in note sequences, eliminating reliance on external annotations; (2) a lightweight, learnable adapter that bridges two pre-trained language model encoders, enabling parameter-efficient, task-agnostic modeling; and (3) a symbolic-music-specific tokenization scheme integrated with functional semantic alignment. The framework achieves state-of-the-art performance across multiple symbolic music understanding and generation benchmarks. All code, pretrained models, and interactive demos are publicly released.

This study addresses the unclear effectiveness of adapting large language models (LLMs) to symbolic music tasks and the lack of systematic comparisons among fine-tuning strategies. We present the first comprehensive evaluation of supervised fine-tuning and preference optimization methods, assessing general-purpose instruction-tuned LLMs, domain-adapted variants, and specialized music LLMs across multiple ABC notation corpora on both generation and understanding tasks. Our experiments reveal a trade-off between effective domain adaptation and preservation of pre-trained knowledge, while also uncovering inconsistencies in commonly used automatic evaluation metrics within the symbolic music context. These findings offer empirical insights and practical guidance for future research on LLM adaptation in symbolic music processing.

This study addresses the lack of lightweight, corpus-driven style imitation tools for symbolic music composition. We propose a corpus-based symbolic music style imitation method and implement it in Apollo, an interactive desktop application. Apollo enables users to construct custom Western classical music corpora; it performs low-latency stylistic melody generation via MIDI parsing and local pattern matching. The system supports corpus management, real-time preview, controllable parameter adjustment, and MIDI export/streaming. Its primary contribution is the first corpus-driven, user-customizable, and real-time interactive symbolic music style generation system designed specifically for musicians and researchers—filling a critical gap in lightweight creative assistance tools. Experimental evaluation demonstrates that Apollo reliably generates stylistically coherent musical phrases and significantly enhances compositional exploration efficiency and stylistic analysis capability.

Controllable symbolic music generation is hindered by the scarcity of large-scale, high-quality datasets annotated with rich metadata (e.g., instrumentation, style, composer) and titles. Method: We introduce MetaScore—the first 963K-sample symbolic music dataset with fine-grained annotations—and propose a novel LLM-augmented data construction paradigm: leveraging large language models to generate pseudo-natural-language descriptions for scores, integrated with REMI/ABC representations and metadata-driven alignment. Based on MetaScore, we train dual-path conditional generative models—diffusion and Transformer—driven jointly by text prompts and predefined categorical labels. Contribution/Results: Our approach achieves statistically significant improvements over baselines in subjective listening evaluations. The text interface enables open-domain natural-language control, while the label-based system supports high-precision structured generation; both attain production-level quality. MetaScore and the proposed framework jointly address the longstanding gaps in high-quality annotated symbolic music data and controllable generation infrastructure.

This work addresses the limited controllability of existing symbolic music generation methods, which typically rely on from-scratch synthesis and struggle to support explicit local editing. The paper introduces the first explicit editing framework for symbolic music, reframing generation as a draft-editing process. Built upon a BEAT-based rhythmic grid anchoring scheme, the framework unifies three editing mechanisms—token-wise sequence labeling, iterative accompaniment refinement, and post-hoc token infilling—within a single pre-trained backbone model to enable efficient inference. Experimental results demonstrate that the proposed approach outperforms both autoregressive and diffusion models across three distinct editing tasks, achieving inference latency under 100 milliseconds while significantly improving both generation accuracy and perceptual audio quality. These findings highlight a strong correlation between music representation design and editing efficacy.

Current evaluations of symbolic music in large language models suffer from a lack of standardized representations, datasets, and metrics. To address this gap, this work proposes LilyBench—the first comprehensive benchmark based on LilyPond—which unifies music generation and understanding tasks within a single symbolic framework and introduces a multidimensional evaluation strategy. Integrating LilyPond notation, MusPy descriptors, Jensen-Shannon similarity, LilyBERT embeddings, and Fréchet Music Distance, the benchmark enables compilable score generation under zero-shot settings. Experimental results demonstrate that existing open-source large models excel at composer and genre identification but still face significant challenges in structural music understanding. The study also reveals systematic discrepancies among different evaluation metrics, highlighting the need for more coherent assessment methodologies in symbolic music modeling.

This study addresses the lack of unified, reproducible information-theoretic tools in symbolic music analysis, particularly for harmonic progressions and rubato modeling. We present vega-mir, an open-source Python toolkit integrating nine quantitative measures—including Shannon entropy, Kullback–Leibler divergence, Zipf’s law fitting, chord transition graph networks, and rubato spectral analysis—to enable systematic large-scale corpus studies. Innovatively combining PageRank with KL divergence, we reveal for the first time a significant correlation (ρ = 0.61) between harmonic centrality and stylistic deviation across 14 MAESTRO composers. Furthermore, applying Higuchi fractal dimension and periodicity spectrum analysis, we challenge the stereotype of Glenn Gould’s “mechanical” timing, demonstrating that his performances exhibit the strongest rubato periodicity among three pianists, with a median dominant period of 66 beats.

Existing symbolic music generation approaches typically rely on large-scale annotated datasets and support only a single composer’s style, making fine-grained and flexible stylistic control challenging. This work proposes a training-free, inference-stage method that guides generation through manipulation of style vectors in a unified latent space, enabling—for the first time—continuous blending and precise control over multiple composers’ styles. The approach is compatible with various symbolic music generation models and, as demonstrated experimentally, effectively produces compositions in target styles while supporting smooth, interpretable style interpolation. This significantly enhances both creative flexibility and user interactivity in symbolic music generation.

This study investigates the occurrence patterns and structural relationships of motivic variants within their local contexts in symbolic music. To this end, it introduces—for the first time—a multi-label conditional random field (CRF)–based probabilistic graphical model that represents motivic transformations as multi-label variables. The model integrates multi-source musical features, including melodic, rhythmic, and harmonic attributes, to uniformly capture contextual dependency patterns in Beethoven’s piano sonatas. This approach not only offers an interpretable account of co-occurrence regularities and stylistic variation structures among motivic transformation families but also establishes the first quantitative analytical framework for symbolic music corpora that supports joint multi-label modeling of compositional practices.