{"id":"281d1563-cff0-430c-9f02-362c20099748","arxiv_id":"2501.04630","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Interval-based pitch tokenization, where each note is encoded by its distance to a chosen reference, improves Transformer performance and interpretability on several symbolic music analysis tasks.","lead":"This paper tests replacing absolute MIDI pitch numbers in symbolic music tokens with interval distances to a chosen reference line, such as the melody, the top voice, or the bass. In three music-analysis tasks, the interval-based encodings often outperform the absolute encoding, and the best reference line matches musical intuition for each task.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Post-hoc selection of the best intervalization reference in Fig. 2, without multiple-comparison or significance testing, makes the 'improves all tasks' claim unsupported; the era-classification gain is only 1.2% and reference choice has no effect.","rationale":"I agree with the reader that the homophonic restriction limits the scope of the empirical claim, but I see the more load-bearing issue in the inference from the reported results: the headline improvement is obtained by selecting the best intervalization reference after seeing the results, with no correction for multiple comparisons and no variance reporting. The reader's rationale mentions the post-hoc selection and missing significance testing, so we partially agree, but the reader's stated weakest assumption is the homophonic-texture restriction. I would elevate the statistical-selection issue because it affects the central claim even within the homophonic datasets and because the marginal era-classification result suggests the effect may not survive a fair test. The paper deserves credit for releasing code and data, which makes the concrete test feasible, and the interpretability analysis of chord inversions is a genuine strength. My recommendation is to keep the verdict conditional, but to make the condition explicit: demonstrate that a pre-specified or appropriately corrected intervalization comparison yields significant gains on all three tasks.","tokens_in":10627,"tokens_out":4775,"duration_ms":50235,"concrete_test":"Using the released code and models, recompute era classification accuracy for every tokenization in Table 1 across the three seeds, not just the best-reference variant. Then run a paired bootstrap test (10,000 resamples) comparing the maximum intervalized variant accuracy against the absolute-baseline accuracy in the pre-trained setting. If the 95% confidence interval includes zero or a multiple-comparison-adjusted test is nonsignificant, the claim \"Intervalization improves the model performance for all tasks\" should be narrowed to the phrase and chord tasks.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim, \"Intervalization improves the model performance for all tasks, both for pre-trained and end-to-end models,\" rests on Figure 2, whose caption states \"The intervalized model is based on the reference resulting in the best performance.\" This is post-hoc selection over six intervalized variants (three references crossed with two interval-encoding choices) compared against one absolute encoding. Because Figure 3 shows the reference choice is consequential (melody is best in 11/12 comparisons for start-of-phrase, bottom-line in 10/12 for chord inversion), the gains in Figure 2 are maxima over variants rather than a fixed, pre-specified tokenization. Under a null hypothesis where all tokenizations are equivalent, the maximum of six positively correlated variants will systematically beat the single absolute baseline, producing apparent improvement without any real effect. The paper reports no confidence intervals, standard deviations, or significance tests; only three seeds are used. The era-classification result is explicitly described as a \"marginal 1.2%\" gain, and the text says the reference choice has no significant impact for that task, so the \"improves all tasks\" claim is particularly fragile there. The framework, released code, and interpretability analysis are valuable, but the central empirical assertion is not statistically established as reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a general framework for interval-based tokenization of symbolic music, in which absolute pitch tokens are replaced by pitch intervals computed relative to a reference monophonic sequence (melody, skyline, or bottom line). Six intervalization strategies built on the REMI tokenization are defined, and a BERT-style encoder is trained with and without pre-training on three tasks: era classification, start-of-phrase detection, and chord inversion identification. The paper reports that intervalization improves performance on all three tasks (Figure 2), that the best reference varies by task (Figure 3), and that analyses of vertical-interval-token frequencies reveal musically interpretable patterns for chord inversions (Figure 4). Datasets, code, and pre-trained models are released.","tokens_in":10863,"tokens_out":3367,"duration_ms":31315,"significance":"If the central empirical claim holds, the paper contributes a simple, parameter-free modification to symbolic-music tokenization that yields consistent performance gains across several analysis tasks and provides a natural path toward explainability. The formalization of intervalization is clean and appears to contain no derivation errors. The choice of publicly available datasets, the release of code and models, and the use of three seeds per experimental condition are strengths. The interpretability analysis for chord inversion is suggestive and well aligned with music-theoretic expectations. However, the main empirical comparison is weakened by post-hoc selection of the best intervalization variant without statistical correction, so the headline 'intervalization improves all tasks' claim is not established as reported.","major_comments":[{"comment":"The central claim that 'Intervalization improves the model performance for all the tasks' rests on Figure 2, whose caption states that the intervalized model is based on the reference resulting in the best performance. This is post-hoc selection over six intervalized variants (three references crossed with two interval encodings) against a single absolute baseline. Under the null hypothesis that all tokenizations are equally effective, the maximum of six positively correlated variants will tend to beat the single baseline, so the reported gains are maxima over variants rather than the performance of a fixed, pre-specified tokenization. The problem is exacerbated for era classification, where the gain is described as 'marginal 1.2%' and the text says the reference choice has no significant effect. I ask the authors to report the results for each of the six intervalized variants separately (or use an appropriate multiple-comparison correction), include variance or confidence intervals across the three seeds, and state whether the reference choice was made on a validation set or on the same test results shown in Figure 2.","section":"Results, 'Impact of intervalization', Figure 2"},{"comment":"The 'best model' counts in Figure 3 are based on 12 comparisons per task, but the comparisons are not independent (they share the same pre-trained weights and overlapping data splits), and no significance test is applied to the counts. For example, the claim that 'models trained with a melodic reference achieve the best performance in 11 of 12 comparisons' for start-of-phrase detection is presented as evidence of a musically meaningful role of melody, yet a random ordering of reference models could produce such a count with non-negligible probability given only 12 trials. Please provide a statistical test (e.g., sign test or permutation test) for these counts, or temper the interpretation accordingly.","section":"Results, 'Impact of intervalization references', Figure 3"},{"comment":"The evaluation is restricted to datasets with homophonic texture, for which a monophonic reference (melody, skyline, or bottom line) is explicitly present or approximated. While the paper acknowledges this, the abstract and conclusion state the general claim that intervalization improves model performance, and the formalization is presented as a general framework for polyphonic music. The empirical support for intervalization therefore applies only to homophonic music. I recommend stating this restriction in the abstract and conclusion, and ideally adding at least one polyphonic validation or explicitly positioning the contribution as a homophonic-music result.","section":"Evaluation on downstream tasks, 'Downstream tasks'"}],"minor_comments":[{"comment":"Typo: 'bieng composed' should be 'being composed'.","section":"Evaluation on downstream tasks, 'Era classification'"},{"comment":"The figure would benefit from error bars or a display of per-seed variation; without them, the reader cannot assess the stability of the reported gains.","section":"Figure 2"},{"comment":"The relationship between the six intervalized strategies and the two encoding families (REMI-abs.+VPI and REMI-HPI+VPI) is clear in the table, but the text in 'Impact of intervalization references' says 'we compare twice, once for the end-to-end and once for the fine-tuned' for each split, which is confusing because there are two encoding settings per reference. Please rephrase to make the counting transparent (e.g., state explicitly that each of the three references is tested for both an absolute-reference encoding and a horizontal-pitch-interval encoding, yielding six intervalized models).","section":"Table 1 and surrounding text"},{"comment":"The histograms for second and third inversion use very different scales (counts up to ~1000 for root and first, ~20-40 for second and third). Consider normalizing by class size or using a log scale so that the interval distributions are visually comparable across inversion types.","section":"Figure 4"},{"comment":"In the definition of S_j, the condition '(p, t) ≠ eref_j' is redundant because the reference event is already included in the union. This is not an error, but the notation could be simplified.","section":"Intervalization, notation"},{"comment":"Several reference entries have formatting issues, e.g., 'InProceedings of' and 'InInternational Society for Music Information Retrieval Conference' where spacing or capitalization is inconsistent. Please check the reference list for final typesetting.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The post-hoc selection of the best-performing intervalization variant in Figure 2 is the crux of the paper's central claim. I would encourage the editor to ask for a per-variant reporting of results, a multiple-comparison correction or pre-registered reference choice, and confidence intervals across seeds. The paper's framework and released artifacts are valuable, and the issue is fixable within the scope of the manuscript, so I am not recommending rejection. I also note the paper relies heavily on the authors' prior related work (Kermarec, Bigo, and Keller 2022; Le, Bigo, and Keller 2024) as baselines; this is appropriate but the novelty over those works should be made more explicit in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. The framework is a clean generalization of Kermarec et al.'s skyline-only interval tokenization to arbitrary reference lines (melody, skyline, bottom line), and they test it on three tasks with a small BERT, releasing code and data. That part is genuinely useful. The interpretability analysis for chord inversion, where vertical intervals relative to the bass line line up with music-theoretic definitions, is the best part of the paper and worth reading on its own.\n\nThe problem is the headline empirical claim. Figure 2 compares absolute vs. intervalized, but the intervalized curve is the best of six variants (three references crossed with two encoding choices), chosen post hoc. Under a null where all tokenizations are equivalent, the max of six positively correlated variants will beat a single baseline on average. With no confidence intervals, no significance tests, and only three seeds, the statement that intervalization improves all tasks is not supported as reported. The authors themselves call the era-classification gain a marginal 1.2% and note that reference choice has no impact there, which further undercuts the blanket conclusion. The phrase-detection improvement of 6% is more substantial, but we still need a pre-specified comparison or a multiple-comparison correction to believe it.\n\nThe scope is also narrower than the abstract implies: all datasets are homophonic with a single identifiable melodic track. The intervalization construction needs a reference line, so the results do not transfer to general polyphony. The authors acknowledge this in the text, but it means the practical claim is conditional on a musical property that many symbolic datasets do not have.\n\nThat said, the formalization is simple and correct, and the released artifacts make replication straightforward. The task-dependent reference effects (melody best for phrase segmentation, bottom line best for inversion identification) are a solid, novel finding even if the aggregate claim is premature. I would send this to review, but with a demand: redo the comparison with a pre-specified reference or correct for multiple comparisons, and report variance.\n\nReader, take it for what it is: a useful framework and a good case study in tokenization choice, not a definitive demonstration that intervals beat absolute pitches.","headline":"Useful intervalization framework and a nice interpretability case study, but the 'improves all tasks' claim is post-hoc and statistically unsubstantiated.","tokens_in":11372,"tokens_out":2655,"would_cite":false,"duration_ms":25214,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that replacing absolute MIDI pitch tokens with intervals measured from a chosen reference line (melody, skyline, or bottom line) improves era classification, start-of-phrase detection, and chord inversion identification…","keywords":["symbolic music tokenization","pitch intervals","REMI","music information retrieval","BERT","explainability","chord inversion","phrase detection"],"falsifier":"Train both intervalized and absolute-pitch models on a polyphonic dataset with no single reference line, or on homophonic pieces with the reference chosen at random; if intervalization no longer beats absolute pitches when the reference is random or absent, the claimed improvement depends on having a musically meaningful reference. A more direct check would evaluate the best intervalized and absolute models on the same pieces transposed to many keys; if interval tokens do not make performance more stable across keys, the transposition-invariance intuition is unsupported.","tokens_in":10430,"feed_emoji":"🎵","tokens_out":5036,"duration_ms":44554,"temperature":0.7,"pith_summary":"Symbolic music models typically tokenize pitch as absolute MIDI numbers, ignoring the relational structure that music theory emphasizes. This paper introduces intervalization, a framework that re-encodes every pitch as an interval away from a chosen monophonic reference line while keeping the rest of the sequence tokens unchanged. Applying intervalization to a REMI-style tokenizer and a BERT-style encoder, the paper reports improved performance on all three tasks it studies, with gains from about 1.2% to 6%. It also finds that the reference line matters in a musically interpretable way: melody works best for phrase detection, and the bass line works best for chord inversion identification.","feed_headline":"Interval tokens beat absolute pitches in music analysis","feed_subtitle":"A reference line (melody, skyline, or bass) makes pitch encoding relative, improving three analysis tasks and exposing model logic.","key_machinery":"The carrying object is the intervalization operator I, defined by a reference subsequence xref chosen from the note sequence x, a method Iref for encoding reference events, and a method Inon-ref for encoding all other events. In the studied variants, reference notes are encoded either absolutely or as horizontal intervals from the previous reference note, and non-reference notes are encoded as vertical intervals from the simultaneous reference pitch. This operator turns any absolute-pitch tokenization into a relative one and is the single mechanism through which the paper obtains both the performance gains and the explainability results.","core_discovery":"On the paper's own terms, the central discovery is that intervalization, encoding absolute pitches as intervals relative to a reference subsequence, consistently outperforms absolute pitch encoding in the studied tasks, and the choice of reference is not arbitrary. Models using interval tokens beat their absolute-pitch counterparts across all twelve per-task comparisons, with the largest gain in end-to-end start-of-phrase detection and a modest but consistent gain elsewhere. The paper further shows that the best reference tracks musical structure: melodic reference for phrase boundaries, bottom-line reference for chord inversion identification, and no significant difference for era classification. Finally, inspecting the vertical interval tokens produced by the best chord-inversion model reveals distributions that mirror textbook definitions of inversions, making the tokenization directly readable in musical terms.","pith_inferences":["An untested but plausible corollary is that intervalized models are more invariant under key transposition than absolute-pitch models, since intervals are unchanged by uniform transposition; this could be checked by fine-tuning on one key and evaluating on transposed pieces.","The formal framework permits references outside the piece itself, such as a tonal-centre sequence, which could extend the gains to harmonic tasks where no single performed line serves as a natural reference.","Because time tokens are untouched by intervalization, the same reference-based interval encoding could be combined with different time encodings, such as bar-position versus time-shift, to identify the best joint pitch-time representation."],"forward_implications":["Intervalized tokenization can be applied to existing absolute-pitch tokenizers and sequence models without architectural changes, so the reported gains are cheap to reproduce.","Downstream tasks benefit from different references, so practitioners should select the reference based on the musical property being analyzed rather than defaulting to the skyline.","Because vertical interval tokens encode a note's relation to the reference, inspecting their frequency by predicted class gives a direct, musically interpretable view of model behavior.","Interval tokens help both end-to-end and pre-trained models, and pre-trained models remain about 1.2 times better on average across the studied tasks."],"supporting_citations":[{"why":"Defines the REMI tokenization that the paper modifies and uses as the absolute-pitch baseline.","marker":"Huang and Yang (2020)"},{"why":"Provides the MidiBERT-Piano model architecture and the MIDI-score tokenization used as the base model.","marker":"Chou et al. (2024)"},{"why":"Introduced horizontal and vertical pitch-interval tokenization, which this paper generalizes into the intervalization framework.","marker":"Kermarec, Bigo, and Keller (2022)"},{"why":"Establishes the start-of-phrase detection task and the MTC-Piano dataset construction used for pre-training.","marker":"Le, Bigo, and Keller (2024)"},{"why":"Supplies the When-in-Rome dataset used for the chord inversion identification task.","marker":"Gotham et al. (2023a)"},{"why":"Generates the piano accompaniments used to build the phrase-detection dataset.","marker":"Zhao and Xia (2021)"},{"why":"Describes the masked-language-model pre-training objective applied to all seven tokenization variants.","marker":"Devlin et al. (2019)"},{"why":"Motivates the interval representation by showing that melody memory relies on contour and intervals rather than absolute pitches.","marker":"Dowling and Fujitani (1971)"}],"fun_headline_variants":["Interval tokens beat absolute pitch in music analysis","Relative pitch encoding improves symbolic music models","Pitch intervals outperform absolute values in music AI","Interval-based tokenization wins in symbolic music tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that every piece has a single monophonic reference line (melody, skyline, or bottom line) that can be identified and that every note can be aligned to it; for music without such a line, the intervalization cannot be built, so the claim is only established for homophonic textures.","fun_headline_variants_meta":{"raw":{"variants":["Interval tokens beat absolute pitch in music analysis","Relative pitch encoding improves symbolic music models","Pitch intervals outperform absolute values in music AI","Interval-based tokenization wins in symbolic music tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1316,"prompt_tokens":806,"completion_tokens":510,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":422,"completion_tokens_details":{"reasoning_tokens":454}},"tokens_in":422,"tokens_out":510,"duration_ms":4477,"temperature":1.0,"reasoning_tokens":454,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:27:33.999806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train both intervalized and absolute-pitch models on a polyphonic dataset with no single reference line, or on homophonic pieces with the reference chosen at random; if intervalization no longer beats absolute pitches when the reference is random or absent, the claimed improvement depends on having a musically meaningful reference. A more direct check would evaluate the best intervalized and absolute models on the same pieces transposed to many keys; if interval tokens do not make performance more stable across keys, the transposition-invariance intuition is unsupported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the REMI tokenization that the paper modifies and uses as the absolute-pitch baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MidiBERT-Piano model architecture and the MIDI-score tokenization used as the base model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced horizontal and vertical pitch-interval tokenization, which this paper generalizes into the intervalization framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Generates the piano accompaniments used to build the phrase-detection dataset."},{"cited_title":"J.; and Fujitani, D","cited_arxiv_id":null,"evidence_quote":"Motivates the interval representation by showing that melody memory relies on contour and intervals rather than absolute pitches."}],"review_version":1}