PaperlayerView exact revision on arXiv

Source-linked AI summary

From One-Shot Generation to Incremental Music Composition: Adapting a General-Purpose Instruction LLM for Persistent Symbolic Editing

André Ricardo Ducca Fernandes, Jean-Pierre Briot, Simone Diniz Junqueira Barbosa1, Hélio Côrtes Vieira Lopes

arXiv:2609.34994v1cs.AIcs.LG

TL;DR

Composition often involves successive revisions to shared material, while many music-generation systems focus on complete outputs. This paper adapts a general-purpose instruction LLM to perform operation-aware transitions over a persistent ABC score, with explicit preservation requirements. Checker admission and conditional compliance rise substantially after adaptation, supporting feasibility within the tested scope rather than superior musical quality or human–AI co-creativity.

  • Problem

    Composition often involves successive revisions to shared material, motivating the question of whether a general-purpose instruction LLM can reliably sustain operation-aware editing over a persistent score.

    Source: §1 · pp. 1–2

  • Method

    The paper formulates editing as operation-aware state transitions over persistent ABC and adapts a general-purpose instruction LLM to learn the corresponding preservation contract.

    Source: §1 · p. 2, §3.1 · p. 5, §3.4 · p. 6

  • Results

    Checker admission rises from 29.37% to 99.37%, while conditional operation compliance rises from 0.7205 to 0.9798 after LoRA adaptation.

    Source: §8 · p. 14

  • Takeaways & Limitations

    The results support the feasibility of reliable model-side participation in the specified persistent, operation-aware interaction contract.

    Source: §8 · p. 14

    The study is limited to Irish traditional music, five operations, curated prompt templates, and a primarily monophonic lead-sheet representation; it does not establish cross-genre generalization, superior musical quality, or human–AI co-creativity.

    Source: §6 · p. 12

Abstract

from arXiv · show

Most music-generation systems are still framed and evaluated primarily as producers of complete outputs, whereas composition often proceeds through successive revisions to a shared musical artifact. This paper studies a different use of a general-purpose instruction-following large language model: not as a one-shot music generator, but as a reusable operator over an evolving symbolic score. We formulate incremental composition as a sequence of operation-aware state transitions over persistent ABC notation, with explicit requirements on what each operation may change and what it must preserve. The interaction includes two artifact-initialization variants and three editing operations -- chord addition, inpainting, and transposition. We instantiate the formulation by adapting Llama 3.1 8B Instruct with Low-Rank Adaptation (LoRA) on 496,038 operation-aware dialogue records derived from Irish traditional music. The comparison with the unadapted model is used to test the feasibility of learning this interaction contract, not to claim novelty for fine-tuning itself. Across 500 dialogues per model (1,750 attempted output states), checker admission rises from 29.37% to 99.37%, while compliance conditional on admission rises from 0.7205 to 0.9798. Strict eligibility for reference-relative musical-feature analysis increases from 14 to 1,548 outputs, and Longest Common Subsequence analysis does not show a systematic increase in high-overlap sequences relative to held-out baselines under the specified protocol. The results support the technical feasibility of persistent, operation-aware symbolic editing with a general-purpose instruction LLM. They do not establish superior musical quality or human-AI co-creativity, which remain questions for musician-centered evaluation.

1 INTRODUCTION

Composition often involves revising existing material while preserving some decisions and changing others, but the paper asks whether a general-purpose instruction LLM can reliably sustain that process over a persistent score. It frames this as operation-aware state transitions and tests the interaction contract through adaptation and a feasibility comparison.

1 INTRODUCTION

  • Human-centered studies identify transparency and user control as concerns, while recent systems make revision and persistent composition states more explicit.The paper positions its computational question as complementary to those studies and systems.Source: §1 · p. 2
  • The study tests whether a general-purpose instruction-following LLM can learn to transform a persistent musical artifact while preserving operation-specific invariants.The base/adapted comparison is framed as a feasibility experiment, not as a claim that fine-tuning itself is novel.Source: §1 · p. 2
  • The interaction has two artifact-initialization operations and three editing operations: Add chords, Inpainting, and Transpose.Add chords and Transpose operate globally, while Inpainting targets selected bars; each output becomes the artifact for the next turn.Source: §1 · p. 2
  • The paper formulates conversational symbolic composition as operation-specific state transitions over a persistent artifact, separating transformation scope, preservation invariants, and interaction continuity.It also presents a denominator-aware evaluation and an experimental NONOTO integration in a score-oriented workflow.Source: §1 · p. 2

2 RELATED WORK

Related systems vary across several independent design dimensions, including interaction continuity, transformation scope, control modality, and preservation behavior. The paper distinguishes prior work on symbolic revision and conversational composition from its narrower question of whether a general-purpose instruction LLM can learn a persistent, operation-aware editing contract.

2 RELATED WORK

  • The focused design space compares systems across three qualitative projections, whose marker positions are explanatory rather than scores or rankings.The projections concern interaction continuity versus transformation scope, control modality versus operation semantics, and relation to an existing artifact versus preservation contract.Source: §2 · pp. 2–3, Figure 1 · p. 3

2.1 Symbolic Revision and Assisted Composition

  • Symbolic precedents already support local revision and inpainting, while assisted-composition systems provide workflows for lead sheets, multitrack infilling, and structured controls.BeatEdit makes preservation explicit with typed edits, but uses an edit-specific architecture rather than testing whether a general-purpose instruction LLM can learn the contract.Source: §2.1 · pp. 3–4

2.2 Natural Language and Persistent Conversation

  • Prior language-model work includes ABC-based music generation, while MusiChat maintains an active composition state using LLM reasoning, intent routing, and a symbolic music engine.The paper's narrower concern is whether later operations act on the current harmonized artifact while preserving the specified material.Source: §2.2 · p. 4

2.3 A Focused Design Space and Research Gap

  • The identified gap combines ordinary-language operations, an authoritative score carried across turns, observable preservation invariants, and learning through state-transition supervision in a general-purpose instruction LLM.The evaluation separates operation compliance, musical characterization, and training-corpus overlap; Figure 2 depicts the operation-aware transition over persistent ABC.Source: §2.3 · pp. 4–5, Figure 2 · p. 5

3 METHOD

The method represents incremental composition as transitions over one authoritative ABC artifact, with each operation defining what may change and what must remain stable. It uses ABC in a primarily monophonic, lead-sheet-like Irish traditional music setting.

3.1 Persistent Artifact, Operations, and Scope

  • Each editing request is interpreted relative to the current artifact, and the revised artifact becomes the state for the next turn.Add chords preserves melody, Inpainting leaves material outside selected bars unchanged, and Transpose preserves structural and rhythmic relations while changing tonal content.Source: §3.1 · p. 5

3.2 ABC Representation and Musical Domain

  • ABC provides a text representation with headers and chord symbols that can be passed between user and LLM and parsed across turns.Experiments use primarily monophonic melody with optional chord annotations; the IrishMan corpus contains 216,284 tunes, with 214,122 in training and 2,162 in validation, so the study establishes in-domain rather than cross-genre behavior.Source: §3.2 · pp. 5–6, Figure 3 · p. 6

3.3 Operation-Aware Dialogue Dataset

The paper constructs operation-aware dialogue data for composing over a persistent ABC artifact, then adapts Llama 3.1 8B Instruct with LoRA. The dataset combines multiple operation sequences and initialization scenarios with simulated multi-turn dialogues.

3.3 Operation-Aware Dialogue Dataset

  • The dataset exposes five operations: two artifact-creation options, chord addition, inpainting, and transposition.Inpainting replaces one to four parsed bars while copying material outside the target; transposition applies to the current artifact, including chords when present.Source: §3.3 · p. 6
  • Six scenarios combine melody-only and chord-bearing material with assistant-created or user-supplied starting artifacts.Outputs become the current artifact for the next request; 856 template records across 19 prompt types provide language variation.Source: §3.3 · p. 6
  • 496,038 JSONL records contain 3,241,509 user–assistant operation instances, averaging 6.535 operations per non-empty record.Non-empty records contain 1–20 operations.Source: §3.3 · p. 6

3.4 LoRA Adaptation

  • LoRA adaptation freezes Llama 3.1 8B Instruct’s base weights and learns low-rank update matrices; the trained adapter is merged into the base model for these experiments.Training used rank 64, α = 16, dropout 0.1, an 8,192-token maximum sequence length, and two epochs.Source: §3.4 · p. 6

3.5 Evaluation Protocol

The evaluation separates syntax admission and operation reliability from reference-relative musical features and corpus-relative sequence overlap. Each analytical family has distinct gates and populations.

3.5 Evaluation Protocol

  • The protocol evaluates usable ABC and operation compliance, musical characteristics against IrishMan references, and sequence overlap against training material and held-out data.Syntax admission is shared before the three analytical families branch into separate protocols and populations.Source: §3.5 · p. 6, Figure 4 · p. 7
  • Each model is tested on 500 dialogues across seven output states, with 250 attempts per state and 1,750 attempts per model.The two chains cover melody-only and chord-bearing workflows; Create a melody with chords is trained but is not a dedicated evaluation state.Source: §3.5 · pp. 6–7

3.5.1 Admission and Operation Compliance

  • Admission means a response has extractable, parser-admitted ABC that reaches an operation checker; compliance is assessed only among admitted responses.Reporting both avoids treating a high conditional score on a selectively successful subset as end-to-end reliability.Source: §3.5.1 · p. 7

3.5.2 Reference-Relative Musical Characteristics

  • Nine pitch/rhythm comparisons use pairwise distances, Kullback–Leibler divergence, and overlap area; mirror-image KDE correction keeps non-negative distances on valid support.Chord-bearing states also use CTnCTR, while pitch- and chord-in-scale rates use score-level Mann–Whitney tests with Holm correction.Source: §3.5.2 · p. 7

3.5.3 Corpus-Relative Memorization

  • LCS similarity compares candidate-to-training overlap with held-out validation-to-training overlap; flagged rates use copy > 0.8, and results concern overlap rather than plagiarism or intent.Note and chord pitch classes are encoded relative to the resolved tonic while retaining durations and event order.Source: §3.5.3 · pp. 7–8

4 EXPERIMENTAL RESULTS

LoRA adaptation substantially improves admission and conditional operation compliance, while enabling broader reference-relative feature analysis. Under the specified LCS protocol, the adapted model shows no systematic increase in high-overlap sequences relative to held-out baselines.

4.1 Admission and Compliance

  • Checker admission rises from 0.2937 to 0.9937, while conditional operation compliance rises from 0.7205 to 0.9798.The comparison covers 250 attempts per output state.Source: §4.1 · p. 8, Table 1 · p. 9

4.2 Reference-Relative Musical Characteristics

  • Strict feature eligibility rises from 14 of 1,750 base outputs to 1,548 fine-tuned outputs before equalization, allowing all 63 planned Yang–Lerch comparisons for the adapted model.After scenario equalization, 1,471 outputs are used; the base population is too sparse for an equivalent characterization.Source: §4.2 · p. 9
  • Fine-tuned-model overlap area across nine features ranges from 0.7972–0.9730 for melody transposition and 0.9315–0.9853 for inpainting with chords.These values indicate distributional agreement for selected features, not absolute aesthetic quality.Source: §4.2 · p. 9
  • CTnCTR overlap area is 0.9718 for Add Chords, 0.9652 for inpainting with chords, and 0.9696 for transposition with chords.Pitch- and chord-in-scale comparisons are ceiling-heavy: reference and candidate medians are 1.0 in all ten comparisons.Source: §4.2 · p. 9

4.3 Corpus-Relative Memorization

  • All seven adapted-model rates above the strict copy > 0.8 LCS threshold are below their corresponding held-out rates: 15.60% for chord-bearing and 27.60% for melody-only references.The analysis finds no systematic increase in high-overlap sequences relative to held-out baselines under this representation and threshold; it does not prove absence of memorization or plagiarism.Source: §4.3 · p. 9

4.4 Concrete Multi-Turn Example

  • The author-selected dialogue illustrates persistent artifact handoff, while musical-feature and LCS analyses answer separate questions about the admitted population.Table 2 summarizes successive response states for the evaluated dialogue.Source: §4.4 · pp. 9–10

4.1 Admission and Compliance

  • Across the seven evaluated output states, admission and compliance both improve; chord addition, inpainting, and transposition also show operation-specific preservation and requested changes.The fine-tuned model preserves 4,639 of 4,644 compared melody bars during chord addition, preserves 8,886 of 9,043 non-target bars during inpainting, and matches the requested key in 492 of 496 admitted transpositions.Source: §4.1 · pp. 8–9, Figure 6 · p. 10

5 PROTOTYPE INTEGRATION WITH NONOTO

The NONOTO extension demonstrates a natural-language/ABC route for revising, inspecting, and auditioning symbolic music. It is an architectural prototype, not a professional-composition environment or an HCI evaluation.

5 PROTOTYPE INTEGRATION WITH NONOTO

  • The extended NONOTO interface lets users request a compositional operation in natural language, receive revised ABC, inspect notation, and audition the result.The extension adds a natural-language/ABC path to an interface that already supports score interaction, audio playback, and real-time MIDI output.Source: §5 · p. 10
  • The prototype demonstrates a path to an updated, inspectable, playable artifact, but is not presented as a professional composition environment or an HCI evaluation.Figure 9 illustrates the workflow.Source: §5 · p. 10
  • Figure 7 reports population retained at successive evaluation stages as a percentage of 1,750 attempts per model; the strict feature-eligible musical-feature population is measured before equalization.Source: Figure 7 · p. 11
  • Table 2 is captioned as a four-turn evaluated dialogue illustrating persistent state transitions.The caption supplies the dialogue's scope but not its specific turn contents.Source: Table 2 · p. 11

6 DISCUSSION

The discussion frames adaptation as a feasibility test of whether a general-purpose instruction model can sustain operation-aware editing over a persistent score. Its evidence supports model-side reliability under the tested setup, not musical quality, abstract understanding, or superiority to other editing approaches.

6 DISCUSSION

  • The adapted model is reliable end-to-end under the tested interaction contract, while the unadapted model is unreliable overall despite conditional competence on some admitted operations.Transposition is the clearest example of the base model's conditional competence; many of its responses fail before checker admission.Source: §6 · p. 11
  • The experiment supports learning operation semantics, persistence, and preservation with a general-purpose model, but does not establish preference over dedicated editors, hybrid systems, or deterministic algorithms.BeatEdit uses explicit symbolic edit mechanisms, while MusiChat uses a hybrid LLM/symbolic architecture.Source: §6 · p. 11
  • Editing examples expose authorized changes and invariants, supporting operation-aware behavior in the tested scenarios rather than abstract musical understanding.The reported gains concern chord insertion, preservation outside inpainting targets, target modification, and transposition.Source: §6 · pp. 11–12
  • Selected musical-feature distributions often overlap IrishMan references, but these formative metrics do not define musical quality.Human listening and musician-centered studies remain necessary to assess coherence, usefulness, playability, stylistic plausibility, perceived control, and the practical value of successive edits.Source: §6 · p. 12
  • Tonic-relative, chord-aware LCS shows no systematic excess of thresholded overlap, but neither low overlap nor non-significant tests establish absence of memorization, plagiarism, authorship, or intent.The measure provides reproducible sequence-overlap characterization against a held-out baseline, not a plagiarism determination.Source: §6 · p. 12
  • An exact symbolic algorithm is preferable when transposition alone is required; here, transposition tests use of one language interface for heterogeneous operations on a persistent artifact.The paper does not argue that LLMs should replace deterministic algorithms or solvers when exactness is available.Source: §6 · p. 12
  • The evidence is limited to Irish traditional music, five operations, curated prompt templates, and a primarily monophonic lead-sheet representation, so it does not establish cross-genre generalization or human–AI co-creativity.The operation set also uses whole-artifact scopes for chord addition and transposition, but selected-bar scope for inpainting.Source: §6 · p. 12

7 FUTURE WORK

Future work proposes broader operation scopes, music domains, and musician-centered evaluation, alongside hybrid language-model and constraint-solving approaches. The paper also identifies richer representations, longer dialogues, and recovery from invalid states as extensions.

7 FUTURE WORK

  • Future systems could decouple operation from transformation scope, enabling requests such as transposing or reharmonizing selected bars.The current prototype applies chord addition and transposition globally but inpaints selected bars; future preservation requirements could reflect both operation and region.Source: §7 · pp. 12–13
  • Domain-specific adapters are a proposed route beyond Irish traditional music, with jazz and bossa nova named as initial targets for ongoing evaluation.LoRA permits fixed base weights with small domain adapters learned offline and selected at runtime; FlowComposer is cited as an interaction-level analogy using a different learning mechanism.Source: §7 · p. 13
  • Musician-centered studies are proposed to evaluate successive-edit usefulness, preservation of intent, perceived control, workflow fit, and authorship or ownership perceptions.Where tasks and outputs are comparable, future work could apply common operation-level measures to BeatEdit and MusiChat.Source: §7 · p. 13
  • A proposed hybrid would pair language-model intent interpretation and material proposals with constraint solving for selected hard properties.The discussion identifies musically designed ABC tokenization aligned with notes, durations, chords, and barlines as a possible prerequisite.Source: §7 · p. 13
  • Further extensions include polyphony, richer harmony, more operations, longer and less templated dialogues, invalid-state recovery, and study of cultural and provenance issues.Source: §7 · p. 13

8 CONCLUSION

The conclusion recasts conversational symbolic music generation as incremental composition over a persistent artifact and reports improved interaction reliability after adaptation. The finding is bounded to the tested domain and does not establish superior musical quality or human–AI co-creativity.

8 CONCLUSION

  • The paper frames the task as reliably interpreting successive requests, changing authorized material, preserving required invariants, and carrying the resulting score into the next turn.Source: §8 · p. 14
  • Under 1,750 attempted output states per model, LoRA adaptation raises checker admission from 29.37% to 99.37% and conditional operation compliance from 0.7205 to 0.9798.The conclusion frames this as a feasibility finding about model-side interaction reliability, not novelty for fine-tuning or aesthetic superiority.Source: §8 · p. 14
  • Reference-relative musical distributions characterize selected properties, while tonic-relative LCS characterizes corpus overlap; neither analysis proves or disproves plagiarism or defines absolute musical quality.NONOTO illustrates an edit–render–listen–revise workflow, while future work includes musician-centered interaction and constraint solving.Source: §8 · p. 14
  • The conclusion identifies future work in generalizing operation scope, testing domain-specific adapters and musician-centered interaction, and investigating structured ABC tokenization with explicit constraint solving.Source: §8 · p. 14

Generative AI Use Disclosure

The authors developed the research and initial presentation without generative-AI assistance, then used ChatGPT for English-language editing and exposition refinement.

Generative AI Use Disclosure

  • The authors developed the research problem, system, experiments, analysis, and initial complete presentation without generative-AI assistance.Source: Generative AI Use Disclosure · p. 14
  • ChatGPT was later used for English-language editing and to refine organization, exposition, and rhetorical clarity across several paper sections.Source: Generative AI Use Disclosure · p. 14
  • The authors reviewed and edited AI-assisted text, verified sources and scientific claims, and take responsibility for the final content.Source: Generative AI Use Disclosure · p. 14
Loading 2609.34994v1…