{"id":"c5951975-02e8-4f30-89ba-da666003a627","arxiv_id":"2607.29086","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Intermediate layers of Jukebox and MusicGen represent pitch as a conical helix whose clarity depends on octave-equivalent harmonics in the input.","lead":"Music foundation models Jukebox and MusicGen appear to arrange pitch information in a twisted, helix-like shape in their internal representations. The paper measures this shape with a fitted-helix score and finds it depends on instrument timbre and octave-related harmonics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Helicality is not shown to discriminate a helix from a smooth curve: the fitted model can collapse to a line, and the paper's own pure-tone score (0.619) exceeds its random threshold (0.371), so the threshold cannot certify helicity.","rationale":"The reader's weakest assumption and my concern overlap almost completely: the Helicality protocol is the load-bearing instrument, and it has not been validated against simpler non-helical curves. I agree that this warrants a conditional verdict, not a rejection, because the paper is transparent, code is released, and the reported helical visualizations for some instruments (e.g., Hammond organ in Fig. 9) could survive a stricter test. However, the pure-tone result is a particularly sharp problem: by the paper's own threshold, a condition the authors describe as non-helical would count as helical. That directly undermines the inference. The proposed check—comparing Helicality against line/spiral baselines and inspecting the fitted radius—will settle whether the central claim survives. I do not think the verdict changes from the reader's CONDITIONAL; the paper should be revised to add these baselines and to resolve the pure-tone threshold issue.","tokens_in":11168,"tokens_out":6649,"duration_ms":75336,"concrete_test":"Use the released code to extract the same per-layer pitch-conditioned representations and evaluate three null models under the identical optimization/projection protocol: (1) a straight line (r(p)=0), (2) a planar circle/spiral (c=0), and (3) a permutation null in which pitch labels are shuffled. Report fitted r0/r_slope alongside Helicality. If the line or plane-spiral null exceeds 0.371 on the pure-tone or ocarina conditions, or if r0/r_slope are near zero at high-Helicality layers, the current metric cannot certify helicity; the claim would need a positive test such as a significant Helicality drop under permutation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests entirely on Helicality (Eq. 2), but the implementation does not establish helical geometry. The 9-parameter fit in Eq. (1) includes radius parameters r0 and r_slope that are optimized; nothing prevents the optimizer from collapsing r(p) to near zero, in which case the 'helix' is a straight line and high inverse MSE only shows that pitch-conditioned representations lie on a smooth one-dimensional curve. Maxing over 10 PCA triples makes this worse. The null is uniform-random 3D vectors, which is not the relevant alternative. The paper itself provides a decisive counterexample: in Table 1 the pure-tone condition reaches layer-max Helicality 0.619, above the Random reference 0.371±0.014 used in Sec. 4.2, yet Sec. 4.3 states pure tones do not form a helix. Thus either the threshold is invalid or Helicality measures smoothness rather than helicity. In addition, Helicality as defined here is not the isomap-based metric of [19], so the claimed link to prior work does not supply validation. Without simpler geometric baselines (line, plane spiral) and a check on fitted radii, 'helical structure' is not supported over 'smooth pitch-ordered curve.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes intermediate representations of two generative music foundation models (Jukebox and MusicGen) by feeding isolated notes and applying PCA to the per-layer token-averaged embeddings. It proposes a 'Helicality' score defined as the inverse mean squared error between the pitch-conditioned 3D embeddings and a fitted 9-parameter parametric helix, maximized over the 10 possible triples of the top five principal components. Based on this score, the authors claim that both models embed pitch in a helical structure, that the helix clarity varies with instrument timbre, and that octave-equivalent harmonics (powers of 2) drive helix formation, as supported by a multiple regression analysis on artificial harmonic-controlled signals.","tokens_in":11499,"tokens_out":3103,"duration_ms":31548,"significance":"If the central geometric claim were properly validated, the paper would be a valuable contribution to interpretability of music foundation models, complementing prior probing studies with a structural account of pitch representation. The paper has notable strengths: it analyzes two state-of-the-art generative MFMs, uses a carefully designed set of 137 artificial harmonic conditions, performs statistical tests for harmonic contributions, reports cross-model consistency, and makes code available. However, the load-bearing Helicality metric is not shown to discriminate helical structure from a merely smooth pitch-ordered curve, and the paper contains an internal contradiction between the random-threshold criterion and the pure-tone result. These issues undermine the primary claim as it stands, though they are addressable with additional baselines and calibration experiments.","major_comments":[{"comment":"The Helicality score is inverse MSE after fitting the 9-parameter helix of Eq. (1), which includes free radius parameters r0 and r_slope, and after selecting the best of 10 PCA projections. Nothing prevents the optimizer from collapsing r(p) to near zero, in which case the fitted model is a straight line and a high score reflects only that pitch-conditioned embeddings lie on a smooth one-dimensional curve. The only null model is uniform-random 3D vectors, which is far less structured than real representation manifolds. No comparison is made to simpler geometric alternatives such as a line, a plane circle, or a non-periodic spiral. The claim that the representations are 'helical' rather than merely smooth and pitch-ordered is therefore not established. Please report fitted radii (e.g., r0 and r_slope distributions) and include synthetic and real-data baselines that can discriminate helix","section":"§3.3, Eq. (2), §4.2"},{"comment":"There is a direct internal contradiction. Section 4.2 defines helical structure by exceeding the Random reference of 0.371±0.014, and on this basis concludes that both MFMs embed pitch in a helix. However, Table 1 shows that the pure-tone condition has layer-max Helicality 0.619, which exceeds the random threshold, yet §4.3 states that 'Signals without harmonics do not form a helix.' Thus either the threshold is invalid for certifying helicity, or Helicality measures something other than helical geometry (e.g., smoothness of the pitch-conditioned trajectory). This issue must be resolved before the central claim can be accepted.","section":"Table 1, §4.2, §4.3"},{"comment":"The paper cites [19] as the source of 'Helicality' and claims to apply it, but the metric defined in Eq. (2) is not the isomap-based Helicality of [19]. The cited metric is based on isometric feature mapping and measures octave equivalence in a different way. Consequently, the prior work does not provide external validation for the new parametric-fit score. The authors should either use the original metric, clearly state that they propose a new metric, or calibrate the new metric on synthetic data with known ground-truth shapes (line, circle, helix, noisy helix) to demonstrate that it is sensitive to helical structure specifically.","section":"§2.2, §3.3"}],"minor_comments":[{"comment":"The figure block for Eq. (1) parameters appears repeated many times in the manuscript text, making the paper unnecessarily long and visually confusing. Please replace with a single occurrence.","section":"Fig. 4"},{"comment":"It is unclear whether the Random reference value 0.371±0.014 is computed as the maximum over the 10 PCA projection triples for each random trial or as a score on a single projection. If the former, the threshold is appropriately calibrated to the selection procedure; if the latter, the chance level under the max operation would be higher. Please clarify.","section":"§4.2"},{"comment":"The Helicality score is the inverse of MSE, so it has dimensions of 1/coordinate^2 and depends on the scale of the embeddings. Consider normalizing or reporting a goodness-of-fit measure that is scale-invariant, especially when comparing across layers and models.","section":"Eq. (2)"},{"comment":"The term 'Helicality' is sometimes spelled 'Helicality' (e.g., in the reference [19] title and in related work). Please ensure consistent spelling throughout.","section":"§3.3 / §4.3"}],"recommendation":"major_revision","confidential_remarks":"The central geometric claim is not currently supported because the proposed Helicality metric has not been validated against simpler smooth-curve alternatives, and the pure-tone result in Table 1 contradicts the random-threshold interpretation in §4.2. However, the paper's harmonic regression analysis and artificial-signal framework are promising, and the issue seems fixable within the manuscript's scope by adding baseline models, checking fitted radii, and recalibrating the threshold. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Pitch: the paper shows something real — pitch-conditioned intermediate representations in Jukebox and MusicGen can be fit with a conical helix, and octave-equivalent harmonics (ratios that are powers of 2) are what make the fit clean. That second result is the more durable contribution. It survives even if you drop the word \"helix.\" But the central claim as stated — \"the MFMs embed pitch in helical structure\" — is not yet supported by the metric they built.\n\nWhat's good: the experimental design is thoughtful. Synthesizing harmonic-controlled test signals and running multiple regression over harmonic conditions is a clean way to attribute which spectral components matter. The finding that powers-of-2 harmonics drive the regularity, with a similar ranking across two different models (r=0.78 for artificial signals), is a genuine empirical pattern. Code is open and there is enough detail to reproduce.\n\nThe soft spot is the metric. Helicality here is inverse MSE after fitting a 9-parameter helix (height, radius, phase) to the best of 10 PCA triples. There is no comparison to a straight line or a plane spiral, so a high score shows the points lie on a smooth pitch-ordered curve, not specifically that they wrap around a circle. The only null model is uniform-random vectors, which is a weak reference. The paper even contains the counterexample: in Table 1, the pure-tone condition has layer-max Helicality 0.619, above the random reference of 0.371, yet the text says pure tones don't form a helix. That contradiction means either the threshold is invalid or the metric is measuring smoothness rather than helicity. Also, they call this \"Helicality\" after [19], but [19] uses an isomap-based method; the parametric fit here is a different quantity, so prior validation doesn't carry over.\n\nThat said, these flaws are fixable. Add a straight-line and a spiral baseline, report fitted radii per condition to confirm they're not collapsing, resolve the pure-tone contradiction, and fix the citation. The harmonic-attribution analysis is largely independent of the absolute metric, so that part holds up. Reframed as \"pitch-conditioned representations form regular, sometimes helical, pitch-ordered manifolds whose regularity depends on octave-equivalent harmonics,\" the evidence is solid.\n\nThis paper is for people working on interpretability of audio and music models. It deserves a serious referee; with the revisions above, it would be a useful contribution. I would send it out.","headline":"A useful empirical finding about pitch geometry in MFMs, but the 'helix' claim needs a better metric and simpler baselines.","tokens_in":11996,"tokens_out":2597,"would_cite":true,"duration_ms":26482,"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":"The paper argues that two generative music foundation models, Jukebox and MusicGen, embed pitch in a helical structure in their internal representations, with the clarity and geometry of the helix depending on timbre and driven by octave-eq","keywords":["pitch helix","music foundation models","interpretability","octave equivalence","Helicality","Jukebox","MusicGen","harmonic structure"],"falsifier":"Compute Helicality on a deliberately straight-line arrangement of the same pitch-conditioned embeddings (e.g., by fitting a straight line instead of a helix) and on a planar spiral; if these baselines match or exceed the helix fit, the claim that the structure is specifically helical collapses.","tokens_in":11031,"feed_emoji":"🎵","tokens_out":3555,"duration_ms":33292,"temperature":0.7,"pith_summary":"The paper tries to establish that music foundation models, specifically Jukebox and MusicGen, do not just encode pitch as a flat frequency value but arrange pitches along a three-dimensional helix in their hidden layers, matching the classic pitch helix from music psychology. The authors show this by feeding isolated notes, extracting per-layer representations, projecting into three dimensions, and measuring how closely the points fit a parametric conical helix. They find the helix appears across instruments and that its clarity depends on the harmonic content of the sound: harmonics whose frequencies are powers of two (octave-equivalent) strengthen the helix, while non-octave harmonics weaken it. A sympathetic reader would care because it suggests these models learn a structured, psychologically meaningful representation of pitch rather than a mere statistical correlation.","feed_headline":"Music AI models embed pitch as a helix","feed_subtitle":"Jukebox and MusicGen arrange pitches along a spiral in their hidden layers, mirroring the classic octave cycle of music perception.","key_machinery":"The central object is the parametric conical helix model—a curve defined by nine parameters (height offset, height slope, radius offset, radius slope, angular frequency, phase, and three orthonormal axes) that the authors fit to pitch-conditioned 3D embeddings extracted from each transformer layer. The Helicality score, defined as the inverse mean squared error of this fit, quantifies how closely the representations wrap into a helix. Around this sits the extraction pipeline: time-averaged intermediate representations per note, PCA to five components, and exhaustive selection of the best of ten three-component projections, plus a multiple-regression analysis on artificial signals with contro","core_discovery":"On the paper's own terms: the intermediate representations of the generative music models Jukebox and MusicGen contain a helical ordering of pitch that reflects octave periodicity. Using PCA projections and a nine-parameter conical helix fit, the authors report that all tested instruments yield Helicality scores above a uniform-random baseline, with organ-like timbres scoring highest. Artificial signals with controlled harmonic sets show that harmonics at frequency ratios that are powers of 2 (2, 4, 8 and 1/2, 1/4) contribute positively to helix clarity, while other harmonics contribute negatively, and a natural spectral decay helps. The authors conclude that MFMs embed pitch in a helical st","pith_inferences":["A natural next step is to test whether the helical geometry is causally used for generation, e.g., by intervening on the helix axis and checking whether generated pitch shifts accordingly.","The relative ranking of instruments across the two models (r=0.59) hints that timbre-driven helix clarity may be a general property of music transformers rather than a quirk of one architecture.","The lack of a non-helical geometric baseline (e.g., straight line or planar spiral) means the current evidence cannot distinguish a true helix from any smooth one-dimensional curve; a comparison against such baselines would be a cheap, decisive test."],"forward_implications":["If the helix is real, pitch is represented as a geometric structure combining pitch height and pitch class, analogous to human octave equivalence.","Helicality may serve as a diagnostic of how well a model has internalized musical structure; deeper layers show clearer helices, consistent with probing results.","The difference in helix clarity across timbres implies that the model's pitch geometry is shaped by acoustic input statistics, not a fixed internal code.","The finding that octave-equivalent harmonics drive the helix suggests models rely on the same spectral relationships as human pitch perception."],"fun_headline_variants":["Pitch forms a helix inside music AI models","Music AI packs pitch into octave spirals","Hidden layers of music AI show pitch helix","Octave helix found in music model embeddings","Music models twist pitch into helical geometry"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole argument rests on the assumption that the Helicality score—how tightly the points fit a best-chosen spiral—really measures a spiral shape and not just any smooth curve through the pitch points.","fun_headline_variants_meta":{"raw":{"variants":["Pitch forms a helix inside music AI models","Music AI packs pitch into octave spirals","Hidden layers of music AI show pitch helix","Octave helix found in music model embeddings","Music models twist pitch into helical geometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000129,"raw_usage":{"total_tokens":885,"prompt_tokens":601,"completion_tokens":284,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":345,"completion_tokens_details":{"reasoning_tokens":217}},"tokens_in":345,"tokens_out":284,"duration_ms":3404,"temperature":1.0,"reasoning_tokens":217,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T13:53:41.307215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute Helicality on a deliberately straight-line arrangement of the same pitch-conditioned embeddings (e.g., by fitting a straight line instead of a helix) and on a planar spiral; if these baselines match or exceed the helix fit, the claim that the structure is specifically helical collapses.","supporting_citations":[],"review_version":1}