Pith. sign in

REVIEW 3 major objections 6 minor 40 references

CompLex: Music Theory Lexicon Constructed by Autonomous Agents for Automatic Music Generation

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A music-theory lexicon built by cooperating LLM agents—37,432 items from nine keywords—improves text-to-music generation when used to expand prompts, the paper reports.

desk verdict Genuinely new lexicon resource and multi-agent pipeline, but the headline gains are confounded by MidiCaps overlap and weak statistics; worth a referee, not yet citable. read the letter →

arxiv 2508.19603 v1 pith:KTMJQOIG submitted 2025-08-27 cs.SD cs.AI

Zhejing Hu , Yan Liu , Gong Chen , Bruce X.B. Yu This is my paper · ORCID
classification cs.SDcs.AI
keywords musictheorylexicontext-to-musicgenerationmulti-agentsystemhallucinationmitigationsymbolicaudiopromptengineeringLLMagents
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that a major bottleneck in AI music generation—training data scarcity—can be relieved at generation time rather than by retraining. To do this it builds CompLex, a music-theory lexicon of 37,432 items organized into categories, properties, and values, produced from just nine category keywords and five prompt templates by a multi-agent LLM algorithm called LexConstructor. When a user prompt is expanded with related lexicon entries, the paper reports higher scores across symbolic generation with Text2MIDI and audio generation with MusicGen and Suno, beating both unenhanced prompts and plain LLM-expanded prompts on objective and human-rated measures. The broader claim is that a compact, structured knowledge base, cheaply constructed and automatically checked against hallucinations, can inject musical expertise into models that were never trained on it.

What carries the argument

The load-bearing structure is the quadruple of category, item, property, and property-value pair. It is what makes unstructured music-theory knowledge addressable: each item in a category is described by named properties with concrete values, and related items across categories can be linked by matching those values. The load-bearing process is LexConstructor, a five-role, two-stage multi-agent pipeline whose QA communication strategy—Supervisor asks, Value Explorers answer and refine, Supervisor consolidates—is the mechanism the paper credits for keeping hallucinated values out of the lexicon. Extraction functions over a MIDI reference dataset supply the item lists, so the LLM is not asked

What would settle it

Run a preregistered replication with many randomly sampled prompts and the same three models, pairing CompLex-enhanced and LLM-enhanced outputs and reporting confidence intervals for the CLAP, mood/genre accuracy, and relevance/overall quality differences; use a larger, more diverse rater pool with agreement checks. If the intervals include zero, or raters show no consistent preference, the central claim fails. A simpler side-check: prompt the retriever with the lexicon's item lists removed but identical prompt templates—if gains persist, the lexicon content is not what drives them.

Watch

Extended reading notes

Core claim

The central discovery is the lexicon-as-prompt-enhancer. CompLex turns music theory into four nested structures—categories, items, properties, and property-value pairs—so that a simple request such as "joyful mood" can be expanded into concrete musical directions: suitable keys, tempos, instruments, time signatures, and chord progressions. LexConstructor creates this lexicon in two stages, outline then content, with five agent roles; in the content stage a Supervisor agent generates questions and delegates them to three Value Explorer agents, who brainstorm and refine answers until consensus, a design the paper shows reduces hallucinated values and redundant items. The paper's central claim

Load-bearing premise

The central improvement claim rests on the evaluation being sensitive enough to detect real differences: objective metrics are reported without significance tests or confidence intervals, and the subjective ratings come from 24 online raters with no check of inter-rater agreement, so if those measurements mostly reflect noise or artifacts the reported gains are not established.

Editorial extensions

If this is right

  • Any text-conditioned music model—training-accessible or black-box, symbolic or audio—could gain musical fidelity by expanding prompts from a CompLex-style lexicon, without retraining or weight updates.
  • A two-stage outline-then-content construction pattern, bootstrapped from a structured reference dataset, is a reusable recipe for building large domain lexicons from a handful of human inputs.
  • QA-based multi-agent consensus is a portable guard against hallucinated facts in structured knowledge construction, not just in music.
  • The immediate next applications named by the authors are algorithmic composition and style transfer, where CompLex can act as a shared vocabulary between a user and a generator.
  • Because the lexicon is a separate retrieval component, it can be updated, localized, or extended to expressive and performance vocabulary without changing the music generation model.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper reports averaged gains over its prompt set but not how consistently they hold per prompt; a reader-visible extension would be to report per-prompt deltas and identify prompt types (broad mood words vs precise technical terms) where the lexicon helps least or most.
  • The same pipeline may transfer to other data-scarce, hallucination-prone knowledge domains; the authors do not test this, and confirming it would require building a lexicon of similar structure in a second domain.
  • An important untested comparison is against a conventionally curated lexicon of the same size; if a textbook-derived lexicon of equal schema performs just as well, the gain would come from structured prompt expansion rather than from multi-agent construction.
  • The non-redundancy curve suggests per-item processing over a reference dataset, rather than debate among agents, may be the decisive factor for scaling; a factorial ablation separating tool-based extraction from QA consensus would isolate that mechanism.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes CompLex, a music-theory lexicon containing 37,432 items across 9 categories, constructed automatically by LexConstructor, a multi-agent LLM pipeline. The lexicon is then used to augment user prompts for text-to-music generation. The authors evaluate CompLex-enhanced prompting against raw prompts and LLM-enhanced prompts on Text2MIDI, MusicGen, and Suno, reporting consistent improvements on compression ratio, CLAP, mood accuracy, genre accuracy, and two subjective metrics. They also evaluate LexConstructor against prompting and multi-agent baselines on completeness, accuracy, non-redundancy, and executability, and ablate the QA communication strategy.

Significance. If the reported improvements are robust, the contribution is meaningful: it demonstrates an autonomous pipeline for constructing a structured music-theory lexicon with modest manual input, and it provides evidence that such a lexicon can improve both symbolic and audio text-to-music models without retraining. The construction statistics are transparent, and the hallucination check based on the equal-temperament relation between note names, MIDI values, and frequencies is a sound validation for the Note category. The main empirical claim, however, is not yet established because of a data-overlap confound and the absence of statistical inference.

major comments (3)
  1. [§5.1, §4.3, Table 1] The central attribution of the gains to music-theory knowledge is threatened by a construction/evaluation overlap. CompLex is built from MidiCaps: the Chord Progression category alone has 36,797 of the 37,432 items, collected via Extract_lexical_item(c, Dref) from MidiCaps (§5.1). Text2MIDI is explicitly trained on MidiCaps (§5.2), and the MA/GA metrics come from the MidiCaps pipeline (Melechovsky et al., 2024). Since the prompt enhancer retrieves CompLex items built from the same corpus, CompLex-enhanced prompts can re-inject exact MidiCaps-derived chord progressions and labels into a model and classifiers that already know that distribution. The paper does not state whether the evaluation prompts are MidiCaps captions, how many prompts were used, or whether any held-out/out-of-corpus condition was run. Without this, the reported improvements may reflect distribution matching rather tha
  2. [§5.3, Table 1] The objective metrics (CR, CLAP, MA, GA) are reported as point estimates with no significance tests, confidence intervals, or effect sizes. Several differences are small in magnitude (e.g., Text2MIDI MA 0.47 vs. 0.41; MusicGen CLAP 0.33 vs. 0.25), and it is unclear whether they exceed run-to-run or prompt-to-prompt noise. The subjective evaluation uses only 24 participants and reports no inter-rater reliability, no statistical test, and no protocol validation. The authors should provide paired significance tests (or bootstrap CIs) for the objective metrics and report chance-adjusted agreement for the subjective ratings before claiming consistent improvements.
  3. [§4.3] The prompt-enhancement procedure is underspecified. The text says the system 'queries the lexicon D to retrieve related items I(cu)', but it does not specify how cu is inferred from an arbitrary user prompt, how many items and properties are selected, how the final prompt is assembled, or how irrelevant retrieved items are filtered. The appendix containing the five prompt templates is referenced but not present in the submitted text. This lack of detail prevents replication and makes it difficult to test whether the improvements come from the lexicon's content or from prompt-length/format differences.
minor comments (6)
  1. [§4.1] The paper states that 9 category keywords are manually provided, but the enumerated list contains only 8 (mood, genre, key, instrument, tempo, time signature, chord, note). The missing item appears to be 'chord progression', which is one of the 9 reported categories.
  2. [Figure 4] The label in the figure reads 'MusicLex' while the paper consistently uses 'CompLex'. Please correct the label.
  3. [§5.3, Table 2] The definitions of Completeness, Non-Redundancy, and Executability are too vague to reproduce. For example, Completeness is 'averaged across all categories' but the ground-truth item set is not defined, and Non-Redundancy is described only as a ratio of unique items. Formulas or a precise protocol are needed.
  4. [Table 1] For Suno, the LLM-Enhanced condition yields GA 0.52, below W/o Enhancement (0.67). The statement that 'LLM-Enhanced generally underperforms' should be nuanced, since this is not uniformly true, though CompLex-Enhanced remains the best.
  5. [References] The surname 'Herremans' is misspelled as 'Herremansn' in the Melechovsky et al. reference.
  6. [General] No statement is provided on code/data release or on whether the appendix is part of the published version. Given the reproducibility concerns, this should be clarified.

Circularity Check

1 steps flagged · score 2.0 of 10

Construction/evaluation overlap with MidiCaps confounds the Text2MIDI and MA/GA evidence, but the derivation is not equation-level circular and external models provide independent support.

  1. other [Sections 5.1–5.3; §4.1 (Item Builder)]
    "Reference Dataset: We select the MidiCaps dataset [Melechovsky et al., 2024] as reference dataset... Chord Progression has 36,797 items... Text2MIDI [Bhandari et al., 2025], a SOTA open-source text-to-symbolic music generation model trained on the MidiCaps dataset... Mood Accuracy (MA): Measures how effectively the generated music reflects the mood indicated in the textual description [Melechovsky et al., 2024]."

    The lexicon content is not first-principles: Abuilder populates I(c) via Extract_lexical_item(c,Dref), so ~98% of CompLex (36,797/37,432 chord progressions) is a direct extraction from MidiCaps. The headline text-to-symbolic test model, Text2MIDI, is trained on MidiCaps, and the MA/GA autometrics are the MidiCaps pipeline's classifiers. Hence a CompLex-enhanced prompt reinserts the model's own training-distribution items into the conditioning text. The reported gain for Text2MIDI is therefore at least partly in-distribution familiarization and cannot, by this design, be uniquely attributed to music-theory knowledge. No held-out/out-of-corpus condition or prompt-source disclosure is provided.

full rationale

No load-bearing self-citation, no fitted parameter renamed as prediction, and no uniqueness theorem imported from the authors' own prior work. The multi-agent construction and the hallucination-mitigation evaluation are checked against external equal-temperament ground truth (pitch/MIDI/frequency), and the lexicon-quality metrics (completeness, accuracy, non-redundancy, executability) are not defined in terms of the downstream generation score. The one substantive concern is the MidiCaps overlap: CompLex's item inventory is extracted from MidiCaps, Text2MIDI is trained on MidiCaps, and MA/GA come from the same MidiCaps pipeline. This makes the Text2MIDI and MA/GA portion of Table 1 vulnerable to a distributional confound rather than a clean test of music-theory knowledge. However, the improvements also appear on MusicGen and Suno, which are not trained on MidiCaps, and on CLAP and subjective metrics that do not depend on MidiCaps-trained classifiers. Those provide independent, non-circular evidence that the lexicon-shaped prompt enrichment has at least some genuine effect. The paper would be strengthened by reporting held-out or out-of-corpus prompts, significance tests, and inter-rater reliability, but these are validity concerns, not logical circularity. Overall, the construction of CompLex is an empirical pipeline with external evaluation components, so the central claim does not reduce by definition to its inputs.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central contribution rests on hand-chosen categories and templates, an unverified assumption that agent consensus implies musical fact, and a reference corpus that also appears in the evaluation pipeline. Very few free numerical parameters are fit; the load-bearing inputs are the LLM backbone, the manual request, and the dataset.

free parameters (3)
  • Number of Value Explorer agents K = 3
    QA strategy uses K=3 explorer agents (Section 5.1); no sensitivity analysis is reported, so the hallucination-mitigation claim depends on this hand-chosen count.
  • Manual category keyword set = 9 keywords (mood, genre, key, instrument, tempo, time signature, chord, note, chord progression)
    The entire lexicon scope is defined by these manually chosen categories; the paper markets the method as derived from just 9 manually input category keywords (Section 4.1).
  • Agent prompt templates = 5 templates not disclosed in the provided manuscript
    Each agent role uses one prompt template, and these are load-bearing for LexConstructor's outputs; they are referenced to an appendix that is absent (Section 5.1).
assumptions (3)
  • domain assumption LLM multi-agent QA consensus produces factually correct music-theory values
    Sections 4.2 and Table 3 treat agreement among GPT-4o agents as hallucination mitigation; external verification is only shown for Note category pitch/frequency, not for the other 8 categories.
  • domain assumption MidiCaps is a representative and sufficient reference corpus for lexical item extraction
    Section 5.1 uses MidiCaps to extract items for all categories; completeness is scored against actual items in music theory, but no independent gold standard is described.
  • domain assumption Human evaluator ratings on relevance and musicality measure generation quality
    Section 5.3 defines REL and OVL scales and uses 24 online respondents; the paper assumes these ratings are valid and generalizable without psychometric validation.
invented entities (1)
  • CompLex lexicon (37,432 items)
    purpose: Structured music-theory knowledge for prompt enrichment in text-to-music generation
    The artifact is central but not released; only a small case study (C#6 note) is shown, and the paper's own accuracy evaluation is limited to note values.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CompLex: Music Theory Lexicon Constructed by Autonomous Agents for Automatic Music Generation." pith.science (2026). https://pith.science/paper/KTMJQOIG

@misc{pith2026250819603,
  author       = {Pith},
  title        = {Pith review of: CompLex: Music Theory Lexicon Constructed by Autonomous Agents for Automatic Music Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KTMJQOIG}},
  note         = {Machine review of arXiv:2508.19603}
}
read the original abstract

Generative artificial intelligence in music has made significant strides, yet it still falls short of the substantial achievements seen in natural language processing, primarily due to the limited availability of music data. Knowledge-informed approaches have been shown to enhance the performance of music generation models, even when only a few pieces of musical knowledge are integrated. This paper seeks to leverage comprehensive music theory in AI-driven music generation tasks, such as algorithmic composition and style transfer, which traditionally require significant manual effort with existing techniques. We introduce a novel automatic music lexicon construction model that generates a lexicon, named CompLex, comprising 37,432 items derived from just 9 manually input category keywords and 5 sentence prompt templates. A new multi-agent algorithm is proposed to automatically detect and mitigate hallucinations. CompLex demonstrates impressive performance improvements across three state-of-the-art text-to-music generation models, encompassing both symbolic and audio-based methods. Furthermore, we evaluate CompLex in terms of completeness, accuracy, non-redundancy, and executability, confirming that it possesses the key characteristics of an effective lexicon.

Figures

Figures reproduced from arXiv: 2508.19603 by the authors.

Figure 1
Figure 1. An illustration of CompLex, which contains 9 lexical categories and 37,432 lexical items, each associated with lexical properties [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the LexConstructor algorithm, which includes two stages: Lexicon Outline Creation and Lexicon Content Generation. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. A case study demonstrating LexConstructor mitigate hal [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (1 more)
Figure 3
Figure 3. Figure 3: Ablation Analysis of the Multi-Agent Model: Our [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 34 canonical work pages

  1. [1]

    Gpt-4 technical report

    [Achiam et al., 2023] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Alt- man, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

    arXiv 2023

  2. [10]

    Automatic Analysis and Influence of Hierarchical Structure on Melody, Rhythm and Harmony in Popular Music

    [Dai et al., 2020] Shuqi Dai, Huan Zhang, and Roger B Dan- nenberg. Automatic analysis and influence of hierarchical structure on melody, rhythm and harmony in popular mu- sic. arXiv preprint arXiv:2010.07518,

    work page Pith review arXiv 2020

  3. [12]

    Agent ai: Surveying the horizons of multimodal in- teraction

    [Durante et al., 2024] Zane Durante, Qiuyuan Huang, Naoki Wake, Ran Gong, Jae Sung Park, Bidipta Sarkar, Rohan Taori, Yusuke Noda, Demetri Terzopoulos, Yejin Choi, et al. Agent ai: Surveying the horizons of multimodal in- teraction. arXiv preprint arXiv:2401.03568,

    arXiv 2024

  4. [13]

    Detecting hallucinations in large language models using semantic entropy

    [Farquhar et al., 2024] Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625–630,

    work page 2024

  5. [14]

    Empowering biomedical discovery with ai agents

    [Gao et al., 2024] Shanghua Gao, Ada Fang, Yepeng Huang, Valentina Giunchiglia, Ayush Noori, Jonathan Richard Schwarz, Yasha Ektefaie, Jovana Kondic, and Marinka Zitnik. Empowering biomedical discovery with ai agents. Cell, 187(22):6125–6151,

    work page 2024

  6. [15]

    Instructme: an instruction guided music edit framework with latent diffusion models

    [Han et al., 2024] Bing Han, Junyu Dai, Weituo Hao, Xinyan He, Dong Guo, Jitong Chen, Yuxuan Wang, Yan- min Qian, and Xuchen Song. Instructme: an instruction guided music edit framework with latent diffusion models. In IJCAI, pages 5835–5843,

    work page 2024

  7. [16]

    The beauty of repetition: an algorithmic composition model with motif-level repetition generator and outline-to-music gen- erator in symbolic music generation

    [Hu et al., 2023] Zhejing Hu, Xiao Ma, Yan Liu, Gong Chen, Yongxu Liu, and Roger B Dannenberg. The beauty of repetition: an algorithmic composition model with motif-level repetition generator and outline-to-music gen- erator in symbolic music generation. IEEE Transactions on Multimedia,

    work page 2023

  8. [17]

    Responding to the call: Exploring automatic music composition using a knowledge-enhanced model

    [Hu et al., 2024] Zhejing Hu, Yan Liu, Gong Chen, Xiao Ma, Shenghua Zhong, and Qianwen Luo. Responding to the call: Exploring automatic music composition using a knowledge-enhanced model. In AAAI, volume 38, pages 521–529,

    work page 2024

Show all 40 references
  1. [18]

    Compose with me: Collaborative music in- painter for symbolic music infilling

    [Hu et al., 2025] Zhejing Hu, Yan Liu, Gong Chen, and Bruce XB Yu. Compose with me: Collaborative music in- painter for symbolic music infilling. In AAAI, volume 39, pages 1327–1335,

    2025

  2. [19]

    Ashraful Islam, Mohammed Eunus Ali, and Md

    [Islam et al., 2024] Md. Ashraful Islam, Mohammed Eunus Ali, and Md. Rizwan Parvez. Mapcoder: Multi-agent code generation for competitive problem solving. In ACL, pages 4912–4944. Association for Computational Linguis- tics,

    2024

  3. [20]

    A survey on deep learning for symbolic music generation: Representations, algorithms, evaluations, and challenges

    [Ji et al., 2023] Shulei Ji, Xinyu Yang, and Jing Luo. A survey on deep learning for symbolic music generation: Representations, algorithms, evaluations, and challenges. ACM Computing Surveys, 56(1):1–39,

    2023

  4. [21]

    Trans- former vae: A hierarchical model for structure-aware and interpretable music representation learning

    [Jiang et al., 2020] Junyan Jiang, Gus G Xia, Dave B Carl- ton, Chris N Anderson, and Ryan H Miyakawa. Trans- former vae: A hierarchical model for structure-aware and interpretable music representation learning. In ICASSP, pages 516–520,

    2020

  5. [22]

    Musical voice separation as link prediction: modeling a musical perception task as a multi-trajectory tracking problem

    [Karystinaios et al., 2023] Emmanouil Karystinaios, Francesco Foscarin, and Gerhard Widmer. Musical voice separation as link prediction: modeling a musical perception task as a multi-trajectory tracking problem. In IJCAI, pages 3866–3874,

    2023

  6. [23]

    The socialai school: Insights from developmental psychology towards artificial socio-cultural agents

    [Kovaˇc et al., 2023] Grgur Kovaˇc, R´emy Portelas, Peter Ford Dominey, and Pierre-Yves Oudeyer. The socialai school: Insights from developmental psychology towards artificial socio-cultural agents. arXiv preprint arXiv:2307.07871 ,

    2023 arXiv

  7. [24]

    Audio- gen: Textually guided audio generation

    [Kreuk et al., 2022] Felix Kreuk, Gabriel Synnaeve, Adam Polyak, Uriel Singer, Alexandre D ´efossez, Jade Copet, Devi Parikh, Yaniv Taigman, and Yossi Adi. Audio- gen: Textually guided audio generation. arXiv preprint arXiv:2209.15352,

    2022 arXiv

  8. [25]

    Meloform: Generating melody with musical form based on expert systems and neural networks

    [Lu et al., 2022] Peiling Lu, Xu Tan, Botao Yu, Tao Qin, Sheng Zhao, and Tie-Yan Liu. Meloform: Generating melody with musical form based on expert systems and neural networks. arXiv preprint arXiv:2208.14345,

    2022 arXiv

  9. [26]

    [Mayo et al., 2024] Katherine Mayo, Nicholas Grabill, and Michael P. Wellman. Fraud risk mitigation in real-time payments: A strategic agent-based analysis. In IJCAI, pages 157–165. ijcai.org,

    2024

  10. [27]

    Music perception, pitch, and the auditory system

    [McDermott and Oxenham, 2008] Josh H McDermott and Andrew J Oxenham. Music perception, pitch, and the auditory system. Current opinion in neurobiology , 18(4):452–463,

    2008

  11. [29]

    Chatdev: Commu- nicative agents for software development

    [Qian et al., 2024] Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, et al. Chatdev: Commu- nicative agents for software development. In ACL, pages 15174–15186,

    2024

  12. [30]

    Learning-based methods for com- paring sequences, with applications to audio-to-midi alignment and matching

    [Raffel, 2016] C Raffel. Learning-based methods for com- paring sequences, with applications to audio-to-midi alignment and matching. 331 Ph. D . PhD thesis, thesis, Columbia University,

    2016

  13. [32]

    Theme transformer: Symbolic music generation with theme-conditioned trans- former

    [Shih et al., 2022] Yi-Jen Shih, Shih-Lun Wu, Frank Zalkow, Meinard Muller, and Yi-Hsuan Yang. Theme transformer: Symbolic music generation with theme-conditioned trans- former. IEEE Transactions on Multimedia,

    2022

  14. [33]

    A survey on image data augmen- tation for deep learning

    [Shorten and Khoshgoftaar, 2019] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmen- tation for deep learning. Journal of big data , 6(1):1–48,

    2019

  15. [34]

    Make a song about anything,

    [Suno, 2024] Suno. Make a song about anything,

    2024

  16. [35]

    [Tatar and Pasquier, 2019] Kıvanc ¸ Tatar and Philippe Pasquier

    Ac- cessed: 2024-11-20. [Tatar and Pasquier, 2019] Kıvanc ¸ Tatar and Philippe Pasquier. Musical agents: A typology and state of the art towards musical metacreation. Journal of New Music Research, 48(1):56–105,

    2024

  17. [36]

    A survey on large language model based autonomous agents

    [Wang et al., 2024] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345,

    2024

  18. [37]

    Chain-of-thought prompting elicits reasoning in large language models

    [Wei et al., 2022] Jason Wei, Xuezhi Wang, Dale Schuur- mans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824– 24837,

    2022

  19. [38]

    Jepoo: highly accurate joint es- timation of pitch, onset and offset for music information retrieval

    [Wei et al., 2023] Haojie Wei, Jun Yuan, Rui Zhang, Yueguo Chen, and Gang Wang. Jepoo: highly accurate joint es- timation of pitch, onset and offset for music information retrieval. In IJCAI, pages 4892–4902,

    2023

  20. [39]

    Popmnet: Generating structured pop music melodies using neural networks

    [Wu et al., 2020] Jian Wu, Xiaoguang Liu, Xiaolin Hu, and Jun Zhu. Popmnet: Generating structured pop music melodies using neural networks. Artificial Intelligence , 286:103303,

    2020

  21. [40]

    Large language models as analogical reasoners

    [Yasunaga et al., 2024] Michihiro Yasunaga, Xinyun Chen, Yujia Li, Panupong Pasupat, Jure Leskovec, Percy Liang, Ed H Chi, and Denny Zhou. Large language models as analogical reasoners. In ICLR, 2024

    2024

  22. [2008]

    Midicaps: A large-scale midi dataset with text captions

    [Melechovsky et al., 2024] Jan Melechovsky, Abhinaba Roy, and Dorien Herremansn. Midicaps: A large-scale midi dataset with text captions. In ISMIR,

    2024

  23. [2016]

    Transform- ers in medical imaging: A survey

    [Shamshad et al., 2023] Fahad Shamshad, Salman Khan, Syed Waqas Zamir, Muhammad Haris Khan, Munawar Hayat, Fahad Shahbaz Khan, and Huazhu Fu. Transform- ers in medical imaging: A survey. Medical Image Analy- sis, 88:102802,

    2023

  24. [2017]

    Ai in finance: challenges, techniques, and opportunities

    [Cao, 2022] Longbing Cao. Ai in finance: challenges, techniques, and opportunities. ACM Computing Surveys (CSUR), 55(3):1–38,

    2022

  25. [2018]

    Training verifiers to solve math word prob- lems

    [Cobbe et al., 2021] Karl Cobbe, Vineet Kosaraju, Moham- mad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word prob- lems. arXiv preprint arXiv:2110.14168,

    2021 arXiv

  26. [2019]

    text2midi: Generating symbolic music from cap- tions

    [Bhandari et al., 2025] Keshav Bhandari, Abhinaba Roy, Kyra Wang, Geeta Puri, Simon Colton, and Dorien Her- remans. text2midi: Generating symbolic music from cap- tions. In AAAI,

    2025

  27. [2020]

    Controllable deep melody gener- ation via hierarchical music structure representation

    [Dai et al., 2021] Shuqi Dai, Zeyu Jin, Celso Gomes, and Roger B Dannenberg. Controllable deep melody gener- ation via hierarchical music structure representation. In ISMIR,

    2021

  28. [2021]

    Simple and controllable music gen- eration

    [Copet et al., 2024] Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, and Alexandre D´efossez. Simple and controllable music gen- eration. NeurIPS, 36,

    2024

  29. [2022]

    Chateval: Towards better llm-based evalua- tors through multi-agent debate

    [Chan et al., 2024] Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. Chateval: Towards better llm-based evalua- tors through multi-agent debate. In ICLR,

    2024

  30. [2023]

    Controlling symbolic music generation based on concept learning from domain knowl- edge

    [Akama, 2019] Taketo Akama. Controlling symbolic music generation based on concept learning from domain knowl- edge. In ISMIR, pages 816–823,

    2019

  31. [2024]

    Modeling temporal tonal relations in polyphonic music through deep networks with a novel image-based representation

    [Chuan and Herremans, 2018] Ching-Hua Chuan and Dorien Herremans. Modeling temporal tonal relations in polyphonic music through deep networks with a novel image-based representation. In AAAI, volume 32,

    2018

  32. [2025]

    Deep learning techniques for mu- sic generation–a survey.arXiv preprint arXiv:1709.01620,

    [Briot et al., 2017] Jean-Pierre Briot, Ga ¨etan Hadjeres, and Franc ¸ois-David Pachet. Deep learning techniques for mu- sic generation–a survey.arXiv preprint arXiv:1709.01620,

    2017 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.