Pith. sign in

REVIEW 4 major objections 4 minor 64 references

Exploring GPT's Ability as a Judge in Music Understanding

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A text-only LLM can judge music annotations: GPT detects errors in beat tracking, chord extraction, and key estimation at rates above random when music is converted to symbolic MIDI and fed through structured prompts.

desk verdict A promising but fragile first benchmark for LLMs as MIR annotation judges; the synthetic-error setup and single-run numbers mean the headline accuracies should be read as upper bounds, not robust evidence of music reasoning. read the letter →

arxiv 2501.13261 v1 pith:72ILAYRC submitted 2025-01-22 cs.IR cs.SDeess.AS

Kun Fang , Ziyu Wang , Gus Xia , Ichiro Fujinaga This is my paper · ORCID
classification cs.IRcs.SDeess.AS
keywords LLMasjudgemusicinformationretrievalsymbolicpromptengineeringerrordetectionbeattrackingchordextractionkeyestimation
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

This paper asks whether a text-only large language model can judge music information retrieval output without any music-specific training. It converts music into symbolic MIDI, deliberately inserts errors into beat, chord, and key annotations, and asks GPT to spot those errors. GPT exceeds random baselines in all three tasks, at 65.20% for beat tracking, 64.80% for chord extraction, and 59.72% for key estimation. The authors also show that adding musical concepts to the prompt tends to improve detection, while masking those concepts changes or degrades performance. This matters because it suggests a training-free route to coupling language-model cognition with music analysis systems.

What carries the argument

The central machinery is a six-component prompt: background and task definition, a description of the relevant music concepts, an explanation of the symbolic input format, step-by-step reasoning techniques, an output format specification, and the data itself. Music is encoded as MIDI notes and annotation labels in JSON, which unifies music and text without adapters or fine-tuning. A concept augmentation scheme adds or masks concepts at the music-attribute, task, and domain levels to test whether GPT's judgments track the musical content of the prompt.

What would settle it

Use the same prompts and metrics but replace the synthetic corruptions with annotations from real beat trackers, chord extractors, and key estimators on the same MAPS and POP909 data. If GPT's detection accuracy on these real outputs falls to the random baseline levels reported in Table 2, the paper's claim that GPT exhibits measurable music understanding would be refuted.

Watch

Extended reading notes

Core claim

The paper reports that GPT, given only symbolic MIDI and textual prompts, identifies deliberately corrupted MIR annotations at rates above a random baseline in all three tasks: 65.20% for beat tracking, 64.80% for chord extraction, and 59.72% for key estimation. It also finds that GPT's error-finding performance broadly rises with the amount of musical concept information the prompt provides, while masking all music concepts degrades performance. The authors interpret these results as evidence that text-only reasoning carries a measurable amount of music understanding and can serve as a training-free judge for MIR tasks.

Load-bearing premise

The central claim rests on the assumption that the artificially inserted beat, chord, and key errors at controlled rates are representative of real MIR annotation errors, so that detecting them is a valid measure of music understanding.

Editorial extensions

If this is right

  • A text-only model can serve as a training-free baseline judge for symbolic MIR tasks, giving future work concrete numbers to beat.
  • Prompt design is a substantive variable: explicit music concepts such as rhythm, chroma, and scale measurably shift GPT's detection performance.
  • Full audio-to-text alignment is not required for a first-pass music judge; symbolic input plus prompt engineering is sufficient for above-chance performance.
  • The same six-component prompt and controlled-error protocol can be applied to other symbolic MIR tasks without retraining the model.
  • Concept masking provides a way to observe what happens when music-specific language is removed, isolating the contribution of music knowledge from generic sequence reasoning.

Reading between the lines

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

  • Editorial inference: Because the injected errors are synthetic, the reported accuracies are likely an optimistic estimate for real MIR errors; real errors can be subtler and less uniformly distributed.
  • Editorial inference: The concept-augmentation results could be used as a diagnostic for whether an LLM actually uses music concepts or just responds to wording; one could vary concept names while holding underlying structure constant.
  • Editorial inference: A production system could combine this judge with a perception network, using LLM explanations as weak labels to improve MIR models without human annotation.
  • Editorial inference: The observed trade-off in beat tracking, where adding a rhythm concept raised correct passes on true positives but lowered detection of false positives and negatives, suggests concept prompts may bias the judge toward specific error types.
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

4 major / 4 minor

Summary. The paper explores whether GPT-3.5 can act as a music judge for three MIR tasks by converting performances into symbolic MIDI-like JSON, injecting synthetic errors into beat, chord, and key annotations, and asking GPT to locate or judge those errors. It proposes a six-component prompt template and a concept-augmentation protocol that either adds or masks music concepts. The experiments report above-random F-scores/weighted scores on all three tasks and claim a positive correlation between concept information and detection accuracy.

Significance. The work is timely and useful as a training-free baseline for LLM-based MIR evaluation: it uses public datasets (MAPS and POP909), makes code available, and defines concrete error-detection metrics. If the findings are robust, the result would justify further study of LLM reasoning as a judge for symbolic MIR outputs. However, the current manuscript does not yet establish the two headline claims: the reported accuracies come from different concept settings, the concept-information trend is not consistent across tasks, and the single-run GPT evaluations are not statistically anchored. The significance is therefore conditional on revision.

major comments (4)
  1. [Abstract, Section 4.3, Table 2] The headline numbers 65.20%, 64.80%, and 59.72% are not all from the Basic Concepts condition: in Table 2 they match the '+Rhythm' row for beat tracking (WS=0.6520), the 'Domain Masking' row for chord extraction (f=0.6480), and the '+Scale' row for key estimation (f=0.5972). The text does not disclose this selection. Please either report a single consistent condition for the headline, or explicitly state that these are best per-task settings with the Basic Concepts results reported in the same table.
  2. [Section 4.3, Table 2] The claim of a positive correlation between detection accuracy and the amount of concept information is not supported by Table 2. In chord extraction, the F-score increases from 0.6207 in Basic Concepts to 0.6376 under Music Attribute Masking, 0.6362 under Task Masking, and 0.6480 under Domain Masking, the opposite of the predicted trend. In key estimation, Basic, +Scale, and Task Masking are almost indistinguishable (0.5965, 0.5972, 0.5960). A monotonic or ordinal relation should be tested, or the claim should be restricted to the beat-tracking task, where the ordering is more consistent.
  3. [Section 4.3] The GPT results are single-run and no temperature or model version is reported, even though the paper itself notes "high randomness in output, sensitivity to prompts, and hallucination." Without repeated sampling or confidence intervals, it is unclear whether the observed advantages over the random baseline (e.g., key estimation Basic F-score 0.5965 vs. random 0.5186) are statistically reliable. Please provide repeated trials, report variance, and perform a significance test or at least a conservative interval comparison.
  4. [Section 4.1, Section 5] The evaluation is carried out exclusively on synthetic, content-independent label errors (random beat insertions/deletions/offsets, random chord-attribute flips, random key replacements). Real MIR errors are typically systematic and context-dependent, so above-chance detection of random flips may overstate judging ability for realistic system outputs. Since the conclusion explicitly defers testing on "real MIR errors instead of synthetic ones," the abstract's "solid ground" claim should be softened, or the paper should include at least one experiment on real MIR outputs.
minor comments (4)
  1. [Section 1, first paragraph] "Aconcept augmentation" should read "A concept augmentation."
  2. [Figure 1, chord prompt] "quality of theet chord" contains a typo; it should be "quality of the chord."
  3. [Abstract and Section 4.2] The abstract uses the word "accuracy," but the evaluation reports F-scores and the weighted WS score; please align the terminology throughout.
  4. [Section 4.3, random baselines] The random baseline for chord and key estimation is described as "detects incorrectness with a probability of 50%," but the table reports weighted precision, recall, and F1; a short derivation or explanation of how the tabulated baseline values follow from this rule would help readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the evaluation is an empirical comparison of GPT outputs against random baselines on public datasets, with no fitted parameter or self-citation forcing the central claim.

full rationale

The paper contains no derivation chain of the kind that can be circular. The central claim—that GPT detects deliberately inserted MIR annotation errors above chance—is established by presenting the model with symbolic inputs and prompts and measuring its raw outputs against a random baseline. No parameter is fitted to the test data, and no quantity called a prediction is constructed from the labels it is asked to predict. The synthetic error injection (Section 4.1) is defined independently of GPT's outputs; the random baselines in Table 2 are simple stochastic policies (e.g., choosing k beat labels at random), not optimised to the data. The authors' citation of POP909 [22] is a normal use of a public dataset and is not load-bearing for the conclusion. The acknowledged limitation that real MIR errors may differ from synthetic ones is an external-validity concern, not circularity, because the evaluation task is not defined in terms of the conclusion. I also note, as a reporting inconsistency rather than a circular step, that the abstract's headline accuracies appear to come from different concept-augmentation conditions (beat 65.20 matches the +Rhythm row, chord 64.80 matches the Domain Masking row, and key 59.72 matches the +Scale row), which a reader should verify against Table 2.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new mathematical entities or fitted constants. It relies on domain assumptions about synthetic errors, prompt equivalence, and output parseability.

assumptions (3)
  • domain assumption Synthetic annotation errors injected at fixed rates are representative of real MIR errors and a valid proxy for music understanding.
    The experimental design creates errors by inserting, deleting, offsetting beats, altering chord attributes, and flipping keys (Sections 3.1 and 4.1). The conclusion extends from these synthetic errors to a general statement about GPT's music understanding, but real MIR errors are acknowledged as future work.
  • domain assumption The concept augmentation variants differ only in musical concept content and not in prompt clarity, length, or grammatical structure.
    Section 3.3 describes masking as removing or blurring music concepts, but no analysis is provided showing that the masked prompts are otherwise equivalent. This is load-bearing for the claimed positive correlation between concept amount and accuracy.
  • domain assumption GPT outputs are reliably parseable and the evaluation metrics fairly reflect error detection performance.
    The model is asked to output JSON-like lists and correctness labels. No parse-failure rate or output validation step is reported in Section 4.2, despite the paper's own note about hallucination in the conclusion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring GPT's Ability as a Judge in Music Understanding." pith.science (2026). https://pith.science/paper/72ILAYRC

@misc{pith2026250113261,
  author       = {Pith},
  title        = {Pith review of: Exploring GPT's Ability as a Judge in Music Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/72ILAYRC}},
  note         = {Machine review of arXiv:2501.13261}
}
read the original abstract

Recent progress in text-based Large Language Models (LLMs) and their extended ability to process multi-modal sensory data have led us to explore their applicability in addressing music information retrieval (MIR) challenges. In this paper, we use a systematic prompt engineering approach for LLMs to solve MIR problems. We convert the music data to symbolic inputs and evaluate LLMs' ability in detecting annotation errors in three key MIR tasks: beat tracking, chord extraction, and key estimation. A concept augmentation method is proposed to evaluate LLMs' music reasoning consistency with the provided music concepts in the prompts. Our experiments tested the MIR capabilities of Generative Pre-trained Transformers (GPT). Results show that GPT has an error detection accuracy of 65.20%, 64.80%, and 59.72% in beat tracking, chord extraction, and key estimation tasks, respectively, all exceeding the random baseline. Moreover, we observe a positive correlation between GPT's error finding accuracy and the amount of concept information provided. The current findings based on symbolic music input provide a solid ground for future LLM-based MIR research.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 47 canonical work pages

  1. [1]

    Exploring GPT's Ability as a Judge in Music Understanding

    INTRODUCTION Recent advancements in text-based Large Language Mod- els (LLMs) have showcased their significant reasoning and knowledge retrieval capabilities across various domains, including music understanding. For instance, the standard GPT-4 model performs better than random on music theory questions [1]. This success raises the question of whether su...

    work page Pith review arXiv 2024

  2. [2]

    We pioneer the integration of MIR problems with text-based LLMs. Our approach utilizes prompt- engineering techniques for MIR error detection and adopts the symbolic music format to unify music and text modality, which does not require additional train- ing

    work page

  3. [3]

    We perform a systematic study on GPT’s abilities as a judge in beat tracking, chord extraction, and key estimation tasks, demonstrating GPT’s capability in solving MIR problems

    work page

  4. [4]

    The proposed methodology sets a baseline for future studies

    We provide a solid ground for LLM-based MIR research. The proposed methodology sets a baseline for future studies

    work page

  5. [5]

    In the computer music domain, the research to combine text and audio LLMs is also popular

    RELATED WORK Recently, the advancements of text-based LLMs [6–8] have expanded beyond textual data, incorporating capabili- ties to interpret information from various other modalities. In the computer music domain, the research to combine text and audio LLMs is also popular. For example, Chat- Musician is a text-based LLM, which focuses mainly on generati...

    work page

  6. [6]

    In Section 3.1, we introduce the task definition and data representations for each task

    METHODOLOGY In this study, we use prompt engineering to evaluate the ca- pabilities of text-based LLMs through three MIR error de- tection tasks: beat tracking, chord extraction, and key esti- mation (as shown in Figure 1). In Section 3.1, we introduce the task definition and data representations for each task. In Section 3.2, we discuss the structure and...

    work page

  7. [7]

    correct” or “incorrect

    offset the timing of one beat, where the offset should be greater than a 70ms tolerance [18]. In beat tracking tasks, error detection is not a binary classification problem per detected beat, because there are false negative predictions (i.e., missed beats error). Therefore, it is crucial to return the beat index range so that both false positive beats an...

    work page

  8. [8]

    Chord Root: indicates the fundamental pitch upon which the chord is built…

    work page

Show all 64 references
  1. [9]

    Quality: indicates the tonal quality…

  2. [10]

    Inversion: indicates the inversion… Let's look at some examples:

  3. [11]

    start”: …, “end

    C:maj/3 - This notation represents… The input music and chords are as follows: {“start”: …, “end”: …, “chord_id”: …, “chord_label”: …, “notes”:[{“onset”: …, “duration”: …, “pitch”: …, “velocity”: …}, …]}, … {“start”: …, “end”: …, “chord_id”: …, “chord_label”: …, “notes”:[{“ons...

  4. [12]

    Identify Pitch Classes: This involves disregarding octave information and focusing only on the pitch names (e.g., C, D, E…

  5. [13]

    Determine Chord Quality: Based on the pitch classes present in each segment, determine the quality of theet chord (e.g., major, minor, dominant seventh…

  6. [14]

    beat_id”: …, “beat_location

    Consider Context: ... … The final answer should be in this format: [<CHORD_ID>, <CHORD_ID>, …] Here is the given MIDI notes and chord predictions for your analysis: … Now I will give you a list of performance MIDI notes and beat locations. Please tell me where the wrong beats ...

  7. [15]

    Rhythmic Stability and Metric Hierarchy: …

  8. [16]

    Expressive Nuances and Tempo Analysis: …

  9. [17]

    start_id

    Harmonic and Melodic Context: … … The final answer should be in this format: [{"start_id": <BEAT ID>, "end_id": <BEAT ID>}, …] Here is the list of performance MIDI notes, beat locations, and their indices: … (Analysis by the model) … Based on the analysis above, the incorrect ...

  10. [18]

    Count the occurrences of each pitch class (C, C#, D, etc.) and determine their relative frequencies

    Calculate Pitch Distribution: Calculate the pitch distribution for all measures. Count the occurrences of each pitch class (C, C#, D, etc.) and determine their relative frequencies

  11. [19]

    Identify Dominant Pitch Classes:

  12. [20]

    Determine Tonic Note:

  13. [21]

    correctness

    Analyze Context: ... … The final answer should be in this format: [{“correctness”: “correct” or “incorrect”}] Here is the given MIDI notes and key prediction for you to analyze: … Model Input Model Output Background & Task Definition Music Concepts Description Input Representa...

  14. [22]

    ENSTDkCl

    EXPERIMENTS We conduct our LLM-based MIR tasks with GPT-3.5. We introduce the datasets in Section 4.1 and the evaluation metrics in Section 4.2. The evaluation results are provided in Section 4.3. 4.1 Datasets We use symbolic performance MIDI dataset for the three proposed tas...

  15. [23]

    correct beats

    missing beats, which are also referred to as true posi- tives, false positives, and false negatives, respectively, in conventional beat tracking tasks [18]. We use TP, FP, and FN to denote these sets of beat positions and I to denote the union of time intervals predicted by an...

  16. [24]

    label_id

    The entry at index 213 is incorrect as it has data features instead of a "label_id" and "label_value"

  17. [25]

    label_id

    The entry at index 214 is incorrect as it has data features instead of a "label_id" and "label_value"

  18. [26]

    label_id

    The entry at index 215 is incorrect as it has data features instead of a “label_id” and “label_value’’ … Model Output Now, let's start the analysis. **Step 1: Extract Rhythm Feature** To calculate the rhythm feature, we need to… | Beat ID | Beat Location | On Beat Notes | Off ...

  19. [27]

    Beat at beat_id 213 (beat_location 126.047): - The note at onset 126.048 (pitch 69) starts after this beat, indicating a potential misalignment… - The beat_location of 126.047 seems incorrect based on the timing of the surrounding notes

  20. [28]

    Beat at beat_id 214 (beat_location 126.616): - The beat_location of 126.616 seems incorrect based on

  21. [29]

    The chord label predicts a D minor chord with the flat third (F) in the bass

    Beat at beat_id 215 (beat_location 127.224): - The beat_location of 127.224 seems incorrect based on… Model Output ### Analysis of Chord Predictions … #### Chord ID: 152 - Chord Label: D:min/b3 - Analysis: The notes in this segment contain pitches 77, 62, 65, 50, and 53. The c...

  22. [30]

    **Chord ID: 152** - **Chord Label: D:min/b3** - **Chroma Feature: {'D': 1, 'F': 1, ' A': 1}** - **Analysis: The chroma feature corresponds to a D minor chord. The chord prediction is incorrect as the bass note is not consistent with the label.** - **Correctness: Incorrect** … ...

  23. [31]

    root:quality/inversion

    **Label ID: 152** - **Label: D:min/b3** - **Analysis:** The label might be incorrect as the feature values do not strongly suggest the presence of the "/b3" component in the label. - **Correctness: Incorrect** … Now I will give you a list of performance MIDI notes (start, end,...

  24. [32]

    **Chroma Feature**: Chroma feature represents the distribution of pitch classes in music, disregarding octave…

  25. [33]

    {label_part_0}:{label_part_1}/{label_part_2}

    **Calculation**: - Extract MIDI note data. - Map MIDI notes to pitch classes. - Count occurrences of each pitch class to create a chroma vector.For example, {'C': 1, 'C#': 0, 'D': 1, 'D#': 0, 'E': 1, 'F': 1, 'F#': 0, 'G': 1, 'G#': 0, ' A': 1, ' A#': 0, 'B': 1} means … … Now I ...

  26. [34]

    **Calculate Pitch Distribution**: Count the occurrences of each pitch class (C, C#, D, etc.) and determine their relative frequencies…

  27. [35]

    Also, I will provide you with a label for the entire data

    **Identify Dominant Pitch Classes**: Identify the pitch classes with the highest frequencies in the pitch distribution… Model Input Now I will give you a list of data. Also, I will provide you with a label for the entire data. Please tell me whether the label prediction is cor...

  28. [36]

    **Calculate Pitch Distribution**: - Measure 1: C# (58), F (54), G# (42)…

  29. [37]

    **Identify Dominant Pitch Classes**: F# (20), D# (17), G (16), F (16)

  30. [38]

    Rhythm” 0.8533 0.1496 0.0968 0.6520 - “Beat Location

    **Determine Tonic Note**: F# appears most frequently and consistently… Model Output The provided data consists of four rows, each with various features denoted as 'feature_0', 'feature_1', and 'feature_2’… Based on the provided data, it appears that the label prediction "Eb:mi...

    2008

  31. [39]

    CONCLUSION AND FUTURE WORK In conclusion, we have proposed a methodology to solve MIR problems with text-based LLMs with prompt engi- neering. We evaluate the performance of GPT-3.5 in er- ror detection across three MIR tasks and find out that GPT’s music reasoning ability in ...

  32. [40]

    ACKNOWLEDGMENTS This research has been supported by the Social Sciences and Humanities Research Council of Canada (SSHRC 895-2022-1004) and the China Scholarship Council

    2022

  33. [41]

    Chatmusi- cian: Understanding and generating music intrinsically with llm,

    R. Yuan, H. Lin, Y . Wang, Z. Tian, S. Wu, T. Shen, G. Zhang, Y . Wu, C. Liu, Z. Zhou et al. , “Chatmusi- cian: Understanding and generating music intrinsically with llm,” arXiv preprint arXiv:2402.16153, 2024

    2024 arXiv

  34. [42]

    Spae: Semantic pyramid autoencoder for multimodal generation with frozen llms,

    L. Yu, Y . Cheng, Z. Wang, V . Kumar, W. Macherey, Y . Huang, D. Ross, I. Essa, Y . Bisk, M.-H. Yanget al., “Spae: Semantic pyramid autoencoder for multimodal generation with frozen llms,” Advances in Neural In- formation Processing Systems, vol. 36, 2024

    2024

  35. [43]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models,

    D. Zhu, J. Chen, X. Shen, X. Li, and M. Elhoseiny, “Minigpt-4: Enhancing vision-language understanding with advanced large language models,” arXiv preprint arXiv:2304.10592, 2023

    2023 arXiv

  36. [44]

    Next- gpt: Any-to-any multimodal llm,

    S. Wu, H. Fei, L. Qu, W. Ji, and T.-S. Chua, “Next- gpt: Any-to-any multimodal llm,” arXiv preprint arXiv:2309.05519, 2023

    2023 arXiv

  37. [45]

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

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou et al. , “Chain-of-thought prompting elicits reasoning in large language models,” Advances in neural information processing systems , vol. 35, pp. 24 824–24 837, 2022

    2022

  38. [46]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Alt- man, S. Anadkat et al., “Gpt-4 technical report,” arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  39. [47]

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,

    W.-L. Chiang, Z. Li, Z. Lin, Y . Sheng, Z. Wu, H. Zhang, L. Zheng, S. Zhuang, Y . Zhuang, J. E. Gonzalez et al. , “Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,” See https://vicuna. lmsys. org (accessed 14 April 2023) , vol. 2, no. 3, p. 6, 2023

    2023

  40. [48]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar et al., “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971 , 2023

    2023 arXiv

  41. [49]

    Simple and control- lable music generation,

    J. Copet, F. Kreuk, I. Gat, T. Remez, D. Kant, G. Syn- naeve, Y . Adi, and A. Défossez, “Simple and control- lable music generation,” in Advances in Neural Infor- mation Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New O...

    2023

  42. [50]

    Content-based controls for music large language modeling,

    L. Lin, G. Xia, J. Jiang, and Y . Zhang, “Content-based controls for music large language modeling,” arXiv preprint arXiv:2310.17162, 2023

    2023 arXiv

  43. [51]

    Music understanding llama: Advancing text-to-music generation with question answering and captioning,

    S. Liu, A. S. Hussain, C. Sun, and Y . Shan, “Music understanding llama: Advancing text-to-music generation with question answering and captioning,” CoRR, vol. abs/2308.11276, 2023. [Online]. Available: https://doi.org/10.48550/arXiv.2308.11276

    2023 doi arXiv

  44. [52]

    Video-llama: An instruction-tuned audio-visual language model for video understanding,

    H. Zhang, X. Li, and L. Bing, “Video-llama: An instruction-tuned audio-visual language model for video understanding,” arXiv preprint arXiv:2306.02858, 2023

    2023 arXiv

  45. [53]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Ka- plan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sas- try, A. Askell et al. , “Language models are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  46. [54]

    Least-to-most prompting enables complex rea- soning in large language models,

    D. Zhou, N. Schärli, L. Hou, J. Wei, N. Scales, X. Wang, D. Schuurmans, C. Cui, O. Bousquet, Q. Le et al., “Least-to-most prompting enables complex rea- soning in large language models,” arXiv preprint arXiv:2205.10625, 2022

    2022 arXiv

  47. [55]

    Self- consistency improves chain of thought reasoning in language models,

    X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self- consistency improves chain of thought reasoning in language models,” arXiv preprint arXiv:2203.11171 , 2022

    2022 arXiv

  48. [56]

    Large language mod- els are human-level prompt engineers,

    Y . Zhou, A. I. Muresanu, Z. Han, K. Paster, S. Pitis, H. Chan, and J. Ba, “Large language mod- els are human-level prompt engineers,” arXiv preprint arXiv:2211.01910, 2022

    2022 arXiv

  49. [57]

    React: Synergizing rea- soning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “React: Synergizing rea- soning and acting in language models,” arXiv preprint arXiv:2210.03629, 2022

    2022 arXiv

  50. [58]

    Eval- uation methods for musical audio beat tracking algo- rithms,

    M. E. Davies, N. Degara, and M. D. Plumbley, “Eval- uation methods for musical audio beat tracking algo- rithms,” Queen Mary University of London, Centre for Digital Music, Tech. Rep. C4DM-TR-09-06, 2009

    2009

  51. [59]

    Symbolic representation of musical chords: A pro- posed syntax for text annotations

    C. Harte, M. B. Sandler, S. A. Abdallah, and E. Gómez, “Symbolic representation of musical chords: A pro- posed syntax for text annotations.” in ISMIR, vol. 5, 2005, pp. 66–71

    2005

  52. [60]

    Multipitch esti- mation of piano sounds using a new probabilistic spec- tral smoothness principle,

    V . Emiya, R. Badeau, and B. David, “Multipitch esti- mation of piano sounds using a new probabilistic spec- tral smoothness principle,” IEEE Transactions on Au- dio, Speech, and Language Processing , vol. 18, no. 6, pp. 1643–1654, 2009

    2009

  53. [61]

    A-maps: Augmented maps dataset with rhythm and key annotations,

    A. Ycart, E. Benetos et al., “A-maps: Augmented maps dataset with rhythm and key annotations,” 2018

    2018

  54. [62]

    Pop909: A pop-song dataset for music arrangement generation,

    Z. Wang, K. Chen, J. Jiang, Y . Zhang, M. Xu, S. Dai, X. Gu, and G. Xia, “Pop909: A pop-song dataset for music arrangement generation,” arXiv preprint arXiv:2008.07142, 2020

    2008 arXiv

  55. [63]

    Scikit-learn: Machine learning in python,

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg, J. Vanderplas, A. Pas- sos, D. Cournapeau, M. Brucher, M. Perrot, and Édouard Duchesnay, “Scikit-learn: Machine learning in python,” Journal of Machi...

    2011

  56. [64]

    Survey of hallucination in natural language generation,

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y . Xu, E. Ishii, Y . J. Bang, A. Madotto, and P. Fung, “Survey of hallucination in natural language generation,” ACM Comput. Surv. , vol. 55, no. 12, mar 2023. [Online]. Available: https://doi.org/10.1145/3571730

    2023 doi

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.