What Makes a Good Layer?

Assessing the Layer-Wise Intrinsic Properties of Music Foundation Models

Angelos-Nikolaos Kanatas1, Yuexuan Kong2,3, Pablo Alonso-Jiménez1, Xavier Serra1, Dmitry Bogdanov1. ISMIR 2026.

1Music Technology Group, Universitat Pompeu Fabra; 2Deezer Research; 3Nantes Université, École Centrale Nantes, CNRS, LS2N

Layer-wise downstream performance of 17 music foundation models on 15 tasks. The label-free metrics of each layer and their correlation with performance are on the Metrics page.

Each cell reports the best layer of a model for a task, its score and the profile across layers. Select a cell to open the full curve, or a column header to sort. Headers marked ▸ expand to further metrics of the same datasets. Params counts the evaluated encoder (the audio tower of two-tower models, the decoder of autoregressive models).

ModelParadigmParamsRhythmTimbre
ChordsKeyPitchBeatDownbeatInstrumentEmotionGenreStructureMTG genreMTG instrumentMTG moodMTG top 50MTTSimilarity
MERT-v1-95MMasked95M49.1L1163.4L992.4L187.7L747.5L770.3L466.3L878.6L758.3L416.5L618.5L612.6L727.4L638.4L579.5L2
MuQ (iter)Masked310M66.1L264.5L292.5L275.8L646.5L1279.1L567.1L1085.2L359.9L1117.5L919.7L513.5L1028.3L939.1L279.5L0
MERT-v1-330MMasked330M51.9L265.2L392.2L488.0L947.8L771.4L766.4L1078.3L1057.8L616.2L718.6L1012.6L827.2L937.8L581.0L7
MusicFM (MSD)Masked330M61.6L164.6L490.5L175.7L945.9L1075.5L368.4L584.8L959.4L817.3L918.8L613.6L727.5L739.2L780.5L2
OMAR-RQ (multifeature)Masked580M64.7L267.0L594.2L274.8L1946.4L1473.9L966.8L978.6L1659.4L1616.7L1418.1L913.0L1627.0L1738.4L981.2L14
OMAR-RQ (base)Masked580M50.7L2162.2L290.9L173.7L1843.3L1274.8L1466.0L2181.0L756.7L817.0L918.2L912.9L927.1L1839.7L680.4L3
MERT-v2-30sMasked632M—67.0L2392.3L5——79.9L968.9L891.0L2463.3L1819.5L1721.6L1215.0L1829.8L1540.9L2081.6L7
MERT-v2-FullSongMasked632M—67.0L2393.1L3——79.0L1269.5L790.0L2462.2L2419.6L2321.3L1214.9L2429.7L2240.9L2283.4L7
MusicGen-SAutoregressive300M—68.6L19—85.7L1147.7L14—64.5L1474.8L1458.2L916.1L919.1L914.2L927.0L1038.9L981.3L17
YuE-s1-0.5BAutoregressive0.5B—66.6L24—88.2L1148.8L10—66.7L2181.4L1160.5L916.6L1118.1L1413.1L1026.8L1139.2L882.4L6
MusicGen-MAutoregressive1.5B—65.5L4389.8L085.0L3245.7L2769.4L1753.8L3260.7L2955.4L3813.5L3215.8L4011.2L3223.7L3336.3L2869.3L1
MusicGen-LAutoregressive3.3B—67.2L3990.9L4385.3L3346.5L3369.8L3256.1L3863.8L3556.1L3214.1L3416.3L3911.9L3424.1L3436.3L3268.4L1
YuE-s1-7BAutoregressive7B—63.8L32—85.8L947.4L9—64.1L1674.8L1258.2L1514.4L916.6L1311.8L624.6L3236.5L3280.4L2
MT2Contrastive5.3M61.5L670.0L292.2L9——74.2L855.7L970.3L1251.8L913.9L1017.0L1111.3L1025.0L1034.0L9—
Myna-BaseContrastive22M—53.0L11———70.3L943.2L955.5L1249.6L1211.3L6—9.4L5—31.5L665.5L12
LAION-CLAPContrastive73M—35.9L685.0L8——79.8L1867.6L1984.5L2056.8L1115.7L1818.8L2113.2L1926.0L1840.0L2079.2L11
MAESTSupervised87M—26.8L282.2L2——78.5L565.9L887.2L9—16.6L619.1L513.2L527.5L539.8L584.8L6
Tasks and metrics
ColumnDatasetMetric
KeyGiantStepsweighted score, mir_eval 0.8.2
PitchNSynthaccuracy
ChordsChords1217MIREX score, mir_eval 0.8.2 (root, thirds and triads under the expanded Tonal header)
Beat, DownbeatGTZANF-measure, 70 ms tolerance
InstrumentNSynthinstrument family accuracy
GenreGTZANaccuracy
StructureHarmonixclip-level section accuracy
EmotionEmomusicR² (arousal and valence under the expanded Semantic header)
MTTMagnaTagATune, 50 tagsmacro-averaged AP (macro AUROC under the expanded Semantic header)
MTG genre, instrument, mood, top 50MTG-Jamendo, 87, 40, 56 and 50 tagsmacro-averaged AP (macro AUROC under the expanded Semantic header)
SimilarityDimSimtriplet accuracy with centered cosine similarity (cosine and L2 under the expanded Similarity header)

Every column, including the expanded ones, can be opened in Layer profiles.

Each layer of the frozen encoder is evaluated separately with a trained probe, an MLP with one hidden layer of 512 units or a linear probe (we will extend this to further probe architectures, such as attentive probes). Scores may differ from those reported in the original papers, which use other probes, checkpoints or splits. Per-layer results are available in the data folder.

Citation

@inproceedings{kanatas2026goodlayer,
  title     = {What Makes a Good Layer? Assessing the Layer-Wise Intrinsic Properties of Music Foundation Models},
  author    = {Kanatas, Angelos-Nikolaos and Kong, Yuexuan and Alonso-Jim{\'e}nez, Pablo and Serra, Xavier and Bogdanov, Dmitry},
  booktitle = {Proceedings of the International Society for Music Information Retrieval Conference (ISMIR)},
  year      = {2026}
}

We continue to update these results with more models and downstream tasks.