Retrieval-Augmented Symbolic Music Generation: Fusing Musical Memory with LLM Creativity

Imagine a music generator that can “remember” and remix pieces from a vast musical library, just like how LLMs can look things up when writing text. You’d build a model that retrieves relevant symbolic music snippets to guide new compositions, hypothesizing this improves coherence and stylistic diversity compared to generation from scratch.

Research Question: Can retrieval-augmented generation (RAG) architectures, analogous to those in NLP, enhance the coherence, style-adaptivity, and factual grounding of symbolic music generation?

Hypothesis: Integrating a retrieval module that fetches musical motifs, patterns, or chord progressions from a large symbolic database will help LLM-based music generators produce more stylistically-aware and coherent compositions, as measured by both objective metrics (e.g., motif reuse, tonal consistency) and human evaluators.

Experiment Plan: Design a retrieval-augmented symbolic music generation pipeline: the generator queries a vector database of symbolic music fragments (e.g., themes, motifs, chord progressions) using context from the ongoing composition. Utilize a Transformer-based symbolic music LLM (e.g., a modified MusicBERT or MuseCoco) paired with a retrieval system similar to HippoRAG 2. Collect a large symbolic music dataset (e.g., MAESTRO, Lakh MIDI). Compare outputs to non-retrieval baselines on measures of stylistic coherence, motif integration, and subjective musicality (via listening tests). Analyze whether retrieval improves long-term structure, thematic development, and genre conformity.

References:

• Huang, Y., & Huang, J. X. (2024). A Survey on Retrieval-Augmented Text Generation for Large Language Models. arXiv.org.
• Jim'enez Guti'errez, B., Shu, Y., Qi, W., Zhou, S., & Su, Y. (2025). From RAG to Memory: Non-Parametric Continual Learning for Large Language Models. International Conference on Machine Learning.
• Le, D.-V.-T., Bigo, L., Keller, M., & Herremans, D. (2024). Natural Language Processing Methods for Symbolic Music Generation and Information Retrieval: A Survey. ACM Computing Surveys.