{"id":"e7fae678-c898-4678-b149-7013047784f5","arxiv_id":"2607.19776","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RPPNet generates melodies by planning variable-length perceptually grouped rhythm-pitch primitives first and then decoding them into notes, beating bar-level baselines in subjective structure and musicality ratings.","lead":"RPPNet generates melodies in two stages: it first plans variable-length rhythm-pitch chunks meant to match how listeners group musical phrases, then turns those chunks into notes. In listening tests, raters preferred its melodies over two bar-based generation systems, particularly for long-term structure.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Weight function W(·) is unspecified and unvalidated; without human-boundary agreement the perceptual-grouping claim is unsupported.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: W(·) is unspecified and perceptually unvalidated. I agree fully. My attack sharpens this by noting that the random-grouping ablation does not rescue the perceptual claim, since it only compares to an arbitrary non-perceptual baseline. The paper also provides no internal evidence that W maps to human phrase perception; the only evidence for the perceptual mechanism is indirect (subjective ratings of generated music), which could be confounded by other architectural choices. Therefore the central claim about the psychological representation is not yet established. However, the paper does have some independent support: the ablation demonstrates the heuristic matters, and the subjective results show significant gains over bar-level baselines. These are real empirical contributions. The missing piece is requirement of a perceptual validation and a full W specification. This is addressable, and thus the verdict should remain CONDITIONAL, matching the reader's assessment. No change to the verdict is needed; the concern is valid but not fatal to the entire paper's empirical results.","tokens_in":10959,"tokens_out":2195,"duration_ms":23970,"concrete_test":"Run the released repository's RPP grouping on a corpus with human phrase-boundary annotations (e.g., the Deliege-style annotations or a standard phrase boundary dataset) and compute boundary agreement (e.g., F-measure with a tolerance window or boundary hit rate). If agreement with human annotators is at or above that of existing phrase-boundary detection algorithms, the perceptual claim is supported; if it is near chance or far below, the central attribution collapses. Additionally, in the same test, report the full W(·) specification and a sensitivity analysis by perturbing W's parameters to show that perceptual alignment, not just any heuristic, drives the improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that RPPNet's performance gain stems from the 'structural correctness of the psychological representation' (Section 5.2, Discussion). This hinges entirely on W(·) in Algorithm 1 (Section 3.1), which is never formally specified: the paper only states it scores notes by 'metrical grid position, duration, and syncopation pattern' and defers to the repository. No experiment validates that the resulting RPP boundaries match human phrase perception. The random-grouping ablation (Table 3) shows the heuristic is better than random grouping, but random grouping is an arbitrary baseline; it does not demonstrate that W captures acoustic cues, auditory inertia, and similarity perception as claimed. Without a human-boundary validation, the causal attribution to perception is unsupported — the observed improvement could come from any statistically useful segmentation rule (e.g., duration/contour based) that happens to aid generation. This is the load-bearing weak point because the entire novelty and contribution (Section 1, contribution 2) rest on the perception-driven nature of the representation, not merely on variable-length grouping.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"RPPNet is a two-stage Transformer for monophonic melody generation. The first stage generates a sequence of Rhythm-Pitch Primitives (RPPs), variable-length units of 1–3 notes obtained by a dynamic-programming segmentation heuristic; the second stage decodes each RPP into MIDI notes via a time-scale expansion mapping. The segmentation is claimed to operationalize acoustic cues, auditory inertia, and similarity perception. Subjective listening tests (15 raters, four dimensions) show RPPNet significantly outperforming Museformer and MELONS, while an ablation with randomly grouped RPPs is significantly worse. RPPNet with ground-truth RPPs performs comparably to RPPNet. Objective PPL/SE are reported only for the ablation, not for the external baselines. The paper's causal claim that the performance gain comes from the psychological validity of the RPP grouping depends on the unspecified weight function W(·) and is not validated against human phrase-boundary annotations.","tokens_in":11227,"tokens_out":9005,"duration_ms":81704,"significance":"If the perception-driven grouping claim could be substantiated, the paper would make a useful contribution: it proposes a compact hierarchical representation with explicit variable-length structure, a clean two-stage decoder, and a controlled random-grouping ablation. The subjective comparison is meaningfully designed and the RPPNet-Real condition is a sensible sanity check. However, as submitted, the perceptual foundation is not demonstrated: the manuscript defers the definition of W(·) to a repository and provides no validation against human phrase-boundary annotations. The central novelty therefore rests on an unverified assumption. The paper also ships code on GitHub, which is an asset for reproducibility, but the manuscript text must stand on its own.","major_comments":[{"comment":"The paper's core claim—that the performance gain comes from 'the structural correctness of the psychological representation' (§5.2)—rests entirely on the weight function W(·) used in Algorithm 1. The manuscript does not define W; footnote 2 refers to a repository. It also reports no validation of RPP boundaries against human phrase-boundary annotations. The random-grouping ablation (Table 3) only shows that the heuristic outperforms random segmentation; random segmentation is an arbitrary control and cannot establish that W captures acoustic cues, auditory inertia, or similarity perception. This is load-bearing: if W is arbitrary, the perceptual claim collapses, even if the generated music is preferred. Please (a) specify W completely in the paper or appendix, and (b) validate RPP boundaries against human-annotated phrase boundaries (e.g., GTTM-informed annotations or listener boundary a","section":"§3.1, Algorithm 1, footnote 2"},{"comment":"Objective PPL/SE are reported only for RPPNet versus RPPNet-Random-Grouped. Museformer and MELONS are excluded because PPL requires a consistent vocabulary, but the central comparison in §5.3 therefore rests entirely on one subjective listening test with 15 participants. The abstract and §5.3 claim superiority over baselines; this is not backed by any objective structural measure. The subjective protocol itself is under-specified: no details are given on the number of ratings per melody, the statistical test used, multiple-comparison correction, or inter-rater agreement. Please either report objective metrics for the baselines in a compatible tokenization (or at least structural statistics such as repetition/self-similarity), and/or document the subjective protocol fully so the reader can assess reliability.","section":"§4.3.1, Tables 1 and 3"},{"comment":"The RPP-level token is defined as v_i = {b_i, p_i, d_i, r_i, m_i}, where b_i is the measure index. Eq. (1) factorizes the probability as P(pos_t | h<t) P(d_t, r_t, m_t | h<t, pos_t), omitting b_t. Since note-level decoding uses T_start = b_i R_bar + p_i (§3.3.2), the model must generate b_i or pos_t must encode both bar and metrical position. The notation is ambiguous; if b_i is derived from the cumulative durations of previous RPPs, this should be stated explicitly and Eq. (1) corrected. This is a load-bearing point: the strict alignment claim depends on a well-defined temporal position variable.","section":"§3.2.1, Eq. (1)"},{"comment":"The random-grouping ablation aligns the length distribution, but it does not control for the internal attribute distribution of the resulting RPPs (rhythm-pattern and contour categories). Since W(·) is not specified, the reader cannot determine whether the ablation controls for the perceptual scoring or merely for a particular distribution of RPP types. Please report the attribute distributions of the random-grouped units and, if possible, add an ablation that preserves the marginal distribution of RPP attributes while destroying boundary order, so the effect can be attributed to grouping semantics rather than to unit-type frequencies.","section":"§4.2, RPPNet-Random-Grouped"}],"minor_comments":[{"comment":"These sections contain a nearly identical paragraph on exclusions of MelodyGLM and PhraseLDM; remove the duplicate.","section":"§4.2 and §4.3.2"},{"comment":"SE is called 'Structural Error' in the text but 'structural entropy of Rhythmic Pattern Primitives' in the table footnote; define SE mathematically and use the term consistently.","section":"Table 3, §4.3.1"},{"comment":"'structural quality also approaching non-significance (6.77 vs. 6.51, p=0.08)' should be 'approaching significance' or, more neutrally, 'did not reach significance.'","section":"§5.2"},{"comment":"'Melons' is inconsistent with 'MELONS' used elsewhere; unify capitalization.","section":"Table 1"},{"comment":"The equation uses pos_t while the token definition uses p_i and b_i; choose one notation and map the attributes explicitly.","section":"Eq. (1), §3.2.1"},{"comment":"The claim of being 'the first hierarchical model to do so' should be tempered; PhraseLDM (ref. [36]) also uses phrase-level abstraction. Rephrase to emphasize the perceptually driven grouping rather than any hierarchical unit.","section":"§1"},{"comment":"Please provide training hyperparameters, model dimensions, inference decoding strategy (temperature/sampling), and compute environment in the paper or appendix.","section":"§4.3"},{"comment":"The Pearson correlation over 23 aggregated categories is a weak distributional comparison; report per-category errors or KL divergence as well.","section":"§5.1"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the unspecified W and lack of human-boundary validation make the perceptual claim untestable from the manuscript. Adding an appendix with W and a boundary-agreement experiment is feasible and would materially change the assessment. The paper's architecture and controlled ablation are promising, but the central attribution to perception needs to be supported rather than deferred to a repository."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is worth engaging with, but it has one load-bearing hole. The RPP representation and the strict two-stage decoupling are new relative to Museformer/MELONS, and the random-grouping ablation is the right experiment: it isolates grouping semantics from variable-length capacity. The subjective results, with 15 raters and significance stars, support the claim that RPPNet sounds better than the two bar-based baselines and that random grouping degrades it.\n\nWhat's good: the motivation is grounded in real psychology citations, the grouping idea is specific and computable, and the authors bothered to compare against both a fixed-bar baseline and a random-grouping control. They also publish code. That is more than most papers in this space do.\n\nWhere it breaks: the entire \"perceptual grouping\" contribution rests on W(·), the weight function in Algorithm 1, and the manuscript says only that W scores metrical position, duration, and syncopation, with the full formalization in the repository. That is not enough. The random-grouping ablation shows that some heuristic segmentation works better than random; it does not show that W implements acoustic cues, auditory inertia, and similarity perception. The paper claims the gain \"stems from the structural correctness of the psychological representation,\" but there is no validation of RPP boundaries against human phrase-boundary judgments. Without that, any duration/contour-based rule could produce the same result. This is the central claim, not a side detail.\n\nThe novelty claim \"first hierarchical model to do so\" is slightly overdone—Ref [41] and PhraseLDM already do phrase/segmentation-based hierarchical modeling; the new part is the automatic perceptual grouping, not hierarchy per se. That is fixable by recalibrating the positioning. Also, objective PPL/SE are only reported for the ablation pair, not for Museformer/MELONS, so there is no objective bridge to the baselines; the paper explains the vocabulary mismatch, which is fair, but it means the objective evidence is thinner than it first looks.\n\nThe reader's report and the stress-test note both identify the W gap correctly. I don't see a fabricated flaw here—it is right in the text.\n\nBottom line: for readers working on hierarchical symbolic music generation and structure representation, this is a useful paper to know about, but I would not cite it for the perceptual claim until W is fully specified and validated. It deserves a serious referee, though the referee should demand either a full W specification in the paper or a human-boundary validation (ideally both) before acceptance.","headline":"RPPNet sells a genuinely new variable-length, psychology-flavored structural unit, but the paper never specifies W(·), and that gap keeps the core perceptual claim from being testable.","tokens_in":11670,"tokens_out":2250,"would_cite":false,"duration_ms":24386,"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":"RPPNet claims that replacing fixed bars with perceptually grouped variable-length rhythm-pitch primitives improves the long-term structure and musicality of generated melodies, and that the gain is due to the perceptual grouping, not model","keywords":["melody generation","rhythm-pitch primitive","perceptual grouping","symbolic music","hierarchical Transformer","structure modeling","music psychology","dynamic programming"],"falsifier":"If a listener study were run in which human annotators marked phrase boundaries on a set of melodies and agreement with RPPNet's automatic boundaries was no better than chance, the perceptual grounding claim would be falsified even though the generated melodies might still be preferred. Alternatively, a blind listening test giving randomly grouped variable-length sequences the same length distribution and training budget would, if it scored equal to RPPNet, falsify the claim that the gain stems from perceptual grouping.","tokens_in":10884,"feed_emoji":"🎵","tokens_out":6100,"duration_ms":54755,"temperature":0.7,"pith_summary":"Fixed bars are a poor proxy for how listeners chunk music, so melody generators built on bar-level units tend to produce structurally fragmented long-term output. RPPNet addresses this by generating a sequence of Rhythm-Pitch Primitives (RPPs)—short variable-length groups of at most three notes—whose boundaries are chosen automatically from acoustic cues, auditory inertia, and similarity. A two-stage Transformer first plans the RPP sequence, then expands each primitive into concrete notes via a time-scale expansion mapping. In listening tests, RPPNet outperformed fixed-bar baselines on coherence, rhythmicity, structure, and overall impression, and an ablation with randomly grouped primitives showed the gain comes from the perceptual grouping semantics, not from variable segmentation alone. The residual gap to human-composed music, especially in rhythmic-pattern fidelity, points to where further work is needed.","feed_headline":"Perceptual chunks beat barlines for melody generation","feed_subtitle":"Two-stage model plans variable-length rhythm-pitch primitives, then expands them into notes; listeners rate them higher.","key_machinery":"The Rhythm-Pitch Primitive (RPP) is the paper's central object: a chunk of one to three notes described by note count, a rhythmic pattern (five categories such as iambic and trochaic), and a melodic contour (ascending, descending, undulating). RPP boundaries are formed by a dynamic-programming algorithm that maximizes a structural weight function W(·) over possible chunks, the function encoding metrical position, duration, and syncopation. The architecture then splits generation into an RPP-level Transformer that predicts the global primitive sequence and a Note-level Transformer that expands each primitive into notes, with decoupled serial prediction and a time-scale expansion mapping keepi","core_discovery":"On its own terms, the paper establishes that a hierarchical melody-generation model whose structural units are perceptually grounded variable-length rhythm-pitch primitives—rather than notated bars—produces melodies listeners rate as more coherent, more rhythmic, better structured, and better overall than those of fixed-bar-level baselines. The authors trace the improvement to the grouping rules themselves: replacing them with random groups of the same length distribution degrades every rating dimension significantly.","pith_inferences":["The paper does not specify the weight function W(·) except by reference to the repository, and it reports no validation of RPP boundaries against human phrase annotations; a direct test of boundary agreement would either support or undermine the perceptual-grounding claim.","The same chunking principle may transfer to other temporal sequence domains where fixed tokenization misaligns with perceived units, such as speech prosody or gesture segmentation.","Extending RPPs to polyphonic music would require rethinking the at-most-three-notes constraint, since chords and simultaneous voices do not collapse into monophonic primitives.","A stricter test of the perceptual claim would be a forced-choice listening experiment pitting RPPNet against fixed-bar outputs matched for pitch content, to rule out low-level acoustic differences."],"forward_implications":["Generative models for symbolic music can be built on perceptually motivated variable-length units instead of bars, potentially reducing long-term fragmentation.","The two-stage decoupling (structure first, details second) offers a natural control knob for form and motivic development.","The automatic grouping algorithm can produce phrase-like annotations for training data without human labels.","The ablation indicates structural semantics, not sequence-length variability, drive the quality gain.","The distributional analysis singles out rhythmic patterns as the weakest-learned attribute, suggesting a concrete target for future loss design."],"fun_headline_variants":["Melody generation: perceptual groups beat barlines","Variable-length rhythm-pitch primitives improve melody","Rhythm-pitch chunks align with perception, not barlines","Perceptual primitives yield more musical melodies","Two-stage model plans variable-length primitives for melody"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's central premise is that its dynamic-programming grouping weight function W(·) really captures acoustic cues, auditory inertia, and similarity so that the resulting variable-length units match how human listeners chunk phrases; the paper neither specifies W(·) fully nor validates the boundaries against human annotations.","fun_headline_variants_meta":{"raw":{"variants":["Melody generation: perceptual groups beat barlines","Variable-length rhythm-pitch primitives improve melody","Rhythm-pitch chunks align with perception, not barlines","Perceptual primitives yield more musical melodies","Two-stage model plans variable-length primitives for melody"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000147,"raw_usage":{"total_tokens":976,"prompt_tokens":654,"completion_tokens":322,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":398,"completion_tokens_details":{"reasoning_tokens":246}},"tokens_in":398,"tokens_out":322,"duration_ms":3740,"temperature":1.0,"reasoning_tokens":246,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:43:56.121077+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a listener study were run in which human annotators marked phrase boundaries on a set of melodies and agreement with RPPNet's automatic boundaries was no better than chance, the perceptual grounding claim would be falsified even though the generated melodies might still be preferred. Alternatively, a blind listening test giving randomly grouped variable-length sequences the same length distribution and training budget would, if it scored equal to RPPNet, falsify the claim that the gain stems from perceptual grouping.","supporting_citations":[],"review_version":1}