How Far Can Pretrained LLMs Go in Symbolic Music? Controlled Comparisons of Supervised and Preference-based Adaptation

TL;DR

The paper investigates how far pretrained instruction-tuned LLMs can be adapted to symbolic music using ABC notation, comparing base backbones, supervised fine-tuning (SFT), and preference-based optimization (DPO) against a music-specialized baseline. It employs a unified ABC-style dataset spanning short and long sequence regimes and evaluates generation and understanding via metrics such as perplexity ($\mathrm{PPL}$) and Fréchet Music Distance ($\mathrm{FMD}$), along with MMLU to assess retained general capability. Key findings show that SFT often yields the strongest in-domain understanding improvements, while token-level generation benefits from specialized baselines; DPO can preserve prior capabilities but may hurt token fidelity, and long-form sequences remain harder to adapt. The results highlight important tradeoffs for practical deployment of LLMs in symbolic-music tasks and point to vocabulary-extension and structured prompting as promising future directions.

Abstract

Music often shares notable parallels with language, motivating the use of pretrained large language models (LLMs) for symbolic music understanding and generation. Despite growing interest, the practical effectiveness of adapting instruction-tuned LLMs to symbolic music remains insufficiently characterized. We present a controlled comparative study of finetuning strategies for ABC-based generation and understanding, comparing an off-the-shelf instruction-tuned backbone to domain-adapted variants and a music-specialized LLM baseline. Across multiple symbolic music corpora and evaluation signals, we provide some insights into adaptation choices for symbolic music applications. We highlight the domain adaptation vs.~preserving prior information tradeoff as well as the distinct behaviour of metrics used to measure the domain adaptation for symbolic music.