Selection & Fusion | What Makes a Good Layer?

As published retrospective protocol

Across the paper’s models and tasks: probing only the top-3 proxy-ranked layers comes within 0.4 pp of the best single layer (matching it on 58% of model–task pairs); averaging those three layers lands at 0.3 pp and beats every trainable fusion method (p < 0.001). The per-task proxy metric was chosen retrospectively on the same models (see the paper).

Held out non-circular

Fixing each task family’s proxy on all other models and applying it to a held-out model still places the true best layer in the proxy’s top-3 39.6% of the time (chance 15.1%; top-1: 16.7% vs 5.0%), and 41.7% for tonal tasks via the pitch-transposition equivariance measure.

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B	Myna-Base
Reference
Oracle (best single layer)	63.4	65.2	64.6	64.5	62.2	68.6	65.5	67.2	66.6	47.8
Middle layer	61.1	64.3	64.3	61.9	55.6	66.3	55.2	54.1	54.3	47.8
Last layer	61.0	58.4	62.2	58.0	51.1	63.2	63.4	64.9	66.6	46.7
Proxy-guided fusion (label-free layer choice)
Top-3 avg	61.8-1.6	63.6-1.6	59.5-5.1	64.4-0.1	66.2+4.0	67.8-0.8	65.8+0.3	—	67.0+0.4	45.5-2.3
Top-3 concat	62.8-0.6	63.7-1.5	58.5-6.1	63.3-1.2	65.3+3.1	66.9-1.7	63.5-2.0	—	65.0-1.6	45.1-2.7
Top-5 avg	62.1-1.3	65.2+0.0	60.8-3.8	63.7-0.8	66.6+4.4	67.7-0.9	66.6+1.1	68.2+1.0	67.4+0.8	47.0-0.8
Top-5 concat	60.4-3.0	61.8-3.5	61.3-3.3	61.2-3.3	60.2-2.0	66.4-2.2	63.9-1.6	66.5-0.7	62.5-4.1	46.7-1.1
All-layer fusion (non-trainable)
All-layer avg	64.8+1.4	65.4+0.2	63.5-1.1	64.5+0.0	63.3+1.1	68.0-0.6	63.7-1.8	65.8-1.5	68.0+1.4	46.7-1.1
All-layer concat	59.7-3.7	26.7-38.5	61.0-3.6	63.3-1.2	61.7-0.5	66.0-2.6	62.3-3.2	62.4-4.8	61.8-4.8	40.6-7.2
Trainable fusion
Weighted sum	62.0-1.4	62.9-2.3	62.2-2.3	63.1-1.4	64.8+2.6	66.2-2.4	65.6+0.1	65.6-1.6	64.4-2.2	47.6-0.2
HConv	60.6-2.8	60.1-5.1	61.8-2.8	63.0-1.5	63.8+1.6	62.9-5.7	62.4-3.1	60.6-6.6	61.4-5.2	48.0+0.2
Attentive	57.2-6.2	61.5-3.7	59.8-4.8	61.7-2.8	60.6-1.6	62.0-6.6	58.1-7.4	60.2-7.0	61.7-4.9	39.4-8.4

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-M	MusicGen-L
Reference
Oracle (best single layer)	92.4	92.2	90.5	92.5	90.9	89.8	90.9
Middle layer	90.3	89.9	89.0	90.1	88.1	86.7	85.6
Last layer	90.8	90.8	86.9	87.6	84.4	86.0	88.7
Proxy-guided fusion (label-free layer choice)
All-layer fusion (non-trainable)
All-layer avg	92.4-0.0	90.9-1.3	90.8+0.2	91.9-0.6	94.1+3.2	88.6-1.2	89.3-1.6
Trainable fusion
HConv	90.4-2.0	90.2-2.0	89.7-0.8	91.5-1.0	92.2+1.3	87.3-2.5	87.6-3.3
Attentive	92.8+0.3	91.5-0.7	91.8+1.2	92.8+0.3	94.4+3.5	91.1+1.3	91.6+0.7

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-M	MusicGen-L	Myna-Base
Reference
Oracle (best single layer)	70.3	71.4	75.5	79.1	74.8	69.4	69.8	64.5
Middle layer	70.0	69.4	74.1	78.3	73.5	67.2	68.8	63.0
Last layer	67.7	67.7	72.0	75.2	68.3	64.9	66.7	64.5
Proxy-guided fusion (label-free layer choice)
Top-3 avg	69.6-0.7	68.9-2.5	73.6-1.9	78.8-0.3	72.9-1.9	69.1-0.3	—	63.9-0.6
Top-3 concat	70.5+0.2	69.4-2.0	75.1-0.4	79.7+0.6	74.2-0.6	68.8-0.6	—	62.2-2.3
Top-5 avg	69.0-1.3	68.9-2.5	74.6-0.9	78.4-0.7	72.3-2.5	69.2-0.2	68.7-1.1	63.0-1.5
Top-5 concat	70.7+0.4	70.1-1.3	74.5-1.0	78.5-0.6	73.1-1.7	67.7-1.7	69.0-0.8	59.2-5.3
All-layer fusion (non-trainable)
All-layer avg	69.7-0.6	69.2-2.2	75.5+0.0	78.7-0.4	74.1-0.7	67.7-1.7	69.9+0.1	64.6+0.1
All-layer concat	70.6+0.3	66.5-4.9	76.2+0.7	80.8+1.7	75.1+0.3	69.4+0.0	69.4-0.4	54.7-9.8
Trainable fusion
Weighted sum	69.1-1.2	70.0-1.4	75.2-0.3	81.2+2.2	74.7-0.1	69.9+0.5	69.8+0.0	64.8+0.3
HConv	71.7+1.4	69.2-2.2	74.2-1.3	78.4-0.7	75.2+0.4	67.4-2.0	67.9-1.9	65.9+1.4
Attentive	68.1-2.2	64.6-6.8	73.9-1.6	77.2-1.9	70.0-4.8	69.2-0.2	70.5+0.7	58.6-5.9

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B	Myna-Base
Reference
Oracle (best single layer)	78.6	78.3	84.8	85.2	81.0	74.8	60.7	63.8	81.4	52.1
Middle layer	77.2	74.5	83.1	84.1	74.8	70.7	45.9	43.4	81.4	49.0
Last layer	65.9	72.4	69.7	80.7	59.7	52.4	42.4	44.1	66.6	49.3
Proxy-guided fusion (label-free layer choice)
Top-3 avg	77.6-1.0	75.2-3.1	81.4-3.4	84.8-0.4	80.7-0.3	74.8+0.0	60.7-0.0	60.7-3.1	81.7+0.3	46.5-5.6
Top-3 concat	67.6-11.0	74.5-3.8	83.1-1.7	82.1-3.1	80.7-0.3	72.4-2.4	60.3-0.4	60.0-3.8	80.3-1.1	46.5-5.6
Top-5 avg	77.9-0.7	72.8-5.5	80.7-4.1	84.8-0.4	76.9-4.1	73.8-1.0	60.3-0.4	60.7-3.1	81.7+0.3	46.9-5.2
Top-5 concat	76.9-1.7	73.8-4.5	82.4-2.4	84.1-1.1	79.0-2.0	73.1-1.7	62.4+1.7	56.9-6.9	73.8-7.6	22.4-29.7
All-layer fusion (non-trainable)
All-layer avg	75.5-3.1	70.7-7.6	71.4-13.4	84.1-1.1	76.2-4.8	56.9-17.9	40.3-20.4	40.3-23.5	74.8-6.6	44.5-7.6
All-layer concat	72.8-5.8	67.2-11.1	71.0-13.8	79.0-6.2	70.0-11.0	53.8-21.0	40.0-20.7	45.2-18.6	68.3-13.1	22.8-29.3
Trainable fusion
Weighted sum	66.5-12.0	68.3-10.0	63.1-21.7	70.3-14.9	67.9-13.1	53.5-21.3	38.6-22.1	38.3-25.5	71.7-9.7	46.2-5.9
HConv	67.6-11.0	49.0-29.3	64.1-20.7	80.7-4.5	67.6-13.4	58.6-16.2	37.6-23.1	43.8-20.0	42.1-39.3	49.0-3.1
Attentive	69.7-8.9	64.1-14.2	79.0-5.8	84.5-0.7	77.2-3.8	59.3-15.5	44.8-15.9	42.8-21.0	68.3-13.1	47.9-4.2

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B
Reference
Oracle (best single layer)	87.7	88.0	75.7	75.8	73.7	85.7	85.0	85.3	88.2
Middle layer	87.6	87.7	73.9	75.8	66.0	85.5	84.2	82.5	88.1
Last layer	86.0	87.0	73.5	74.8	55.5	82.6	81.8	81.5	86.5
Proxy-guided fusion (label-free layer choice)
Top-3 avg	87.4-0.3	87.3-0.7	75.5-0.2	74.3-1.5	74.5+0.8	85.3-0.4	84.8-0.2	84.7-0.6	87.7-0.5
Top-3 concat	87.8+0.1	87.6-0.4	76.2+0.5	79.9+4.1	76.8+3.1	85.9+0.2	84.6-0.4	85.0-0.3	87.9-0.3
Top-5 avg	87.3-0.4	87.1-0.9	75.3-0.4	75.0-0.8	75.0+1.3	85.2-0.5	84.5-0.5	84.4-0.9	87.7-0.5
Top-5 concat	87.0-0.7	87.4-0.6	76.5+0.8	79.1+3.3	76.7+3.0	86.1+0.4	84.8-0.2	85.0-0.3	87.6-0.6
All-layer fusion (non-trainable)
All-layer avg	86.1-1.6	87.1-0.9	74.2-1.5	76.9+1.1	74.3+0.6	84.8-0.9	84.0-1.0	84.5-0.8	88.0-0.2
All-layer concat	87.2-0.5	87.2-0.8	79.2+3.5	82.4+6.6	81.6+7.9	85.6-0.1	84.3-0.7	84.6-0.7	88.1-0.1
Trainable fusion
Weighted sum	86.9-0.8	87.3-0.7	74.8-0.9	78.0+2.2	79.4+5.7	85.7+0.0	84.5-0.5	84.7-0.6	88.1-0.1
HConv	87.5-0.2	87.3-0.7	83.8+8.1	84.7+8.9	84.5+10.8	85.5-0.2	84.4-0.6	84.4-0.9	88.0-0.2
Attentive	88.0+0.3	88.6+0.6	80.6+4.9	80.2+4.4	79.6+5.9	86.0+0.3	84.6-0.4	85.1-0.2	88.4+0.2

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B
Reference
Oracle (best single layer)	47.5	47.8	45.9	46.5	43.3	47.7	45.7	46.5	48.8
Middle layer	47.1	46.9	45.0	42.6	43.3	47.5	44.5	43.0	48.7
Last layer	44.5	46.1	44.8	46.5	41.5	40.6	40.5	42.3	46.9
Proxy-guided fusion (label-free layer choice)
Top-3 avg	47.3-0.2	46.9-0.9	44.4-1.5	46.4-0.1	46.0+2.7	47.3-0.4	45.8+0.1	46.1-0.4	48.4-0.4
Top-3 concat	47.4-0.1	47.7-0.1	43.8-2.1	42.6-3.9	44.8+1.5	47.3-0.4	45.2-0.5	46.0-0.5	48.4-0.4
Top-5 avg	47.3-0.2	46.0-1.8	44.4-1.5	45.6-0.9	45.3+2.0	47.1-0.6	45.7-0.0	45.9-0.6	48.3-0.5
Top-5 concat	46.6-0.9	46.5-1.3	44.0-1.9	42.1-4.4	44.2+0.9	47.6-0.1	45.6-0.1	45.7-0.8	48.5-0.3
All-layer fusion (non-trainable)
All-layer avg	46.4-1.1	47.0-0.8	46.0+0.1	44.1-2.4	46.4+3.1	46.3-1.4	44.5-1.2	44.9-1.6	48.5-0.3
All-layer concat	47.2-0.3	47.3-0.5	43.2-2.7	39.5-7.0	40.3-3.0	46.9-0.8	45.0-0.8	45.1-1.4	48.1-0.7
Trainable fusion
Weighted sum	46.8-0.7	46.7-1.1	44.0-1.9	42.1-4.4	41.2-2.1	47.4-0.3	44.8-0.9	46.1-0.4	48.3-0.5
HConv	47.7+0.2	46.8-1.0	42.3-3.6	43.6-2.9	43.8+0.5	47.0-0.7	45.4-0.3	45.1-1.4	48.5-0.3
Attentive	47.1-0.4	47.6-0.2	45.8-0.1	45.0-1.5	46.5+3.2	47.1-0.6	45.4-0.3	45.4-1.1	48.8+0.0

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B	Myna-Base
Reference
Oracle (best single layer)	66.3	66.4	68.4	67.1	66.0	64.5	53.8	56.1	66.7	43.2
Middle layer	65.2	63.7	67.4	67.0	64.2	63.0	50.2	48.3	64.5	41.3
Last layer	49.7	63.0	64.2	64.0	59.3	45.1	31.2	36.6	60.5	43.2
Proxy-guided fusion (label-free layer choice)
Top-3 avg	63.3-3.0	63.7-2.7	64.6-3.8	67.2+0.1	67.7+1.7	64.0-0.5	49.5-4.3	54.0-2.1	67.9+1.2	37.6-5.6
Top-3 concat	63.2-3.1	62.3-4.1	66.2-2.2	70.5+3.4	66.8+0.8	55.2-9.3	44.5-9.3	52.1-4.0	64.5-2.2	41.0-2.2
Top-5 avg	62.7-3.6	64.5-1.9	66.3-2.1	66.4-0.7	68.0+2.0	64.4-0.1	49.5-4.2	54.2-1.9	68.4+1.7	41.2-2.0
Top-5 concat	64.3-2.0	61.1-5.3	66.5-1.9	67.5+0.4	67.6+1.6	63.6-0.9	50.4-3.4	55.9-0.2	64.5-2.2	39.9-3.3
All-layer fusion (non-trainable)
All-layer avg	59.7-6.6	61.2-5.2	62.1-6.3	66.1-1.0	67.5+1.5	55.4-9.1	44.6-9.2	50.4-5.7	66.0-0.7	38.7-4.5
All-layer concat	62.5-3.8	61.2-5.2	65.6-2.8	69.0+2.0	65.6-0.4	61.8-2.7	51.3-2.5	51.7-4.4	65.2-1.5	40.2-3.0
Trainable fusion
Weighted sum	55.7-10.6	62.3-4.1	55.7-12.7	65.0-2.1	61.4-4.6	52.8-11.7	39.4-14.4	50.1-6.0	66.5-0.2	34.0-9.2
HConv	57.1-9.2	61.0-5.5	59.9-8.5	68.9+1.8	59.9-6.1	60.2-4.3	45.1-8.7	50.1-6.0	60.5-6.2	29.8-13.5
Attentive	59.1-7.2	58.1-8.3	58.8-9.6	57.3-9.8	58.5-7.5	50.3-14.2	48.0-5.8	35.1-21.0	55.7-11.0	34.7-8.5

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B	Myna-Base
Reference
Oracle (best single layer)	58.3	57.8	59.4	59.9	56.7	58.2	55.4	56.1	60.5	49.6
Middle layer	57.8	56.1	57.1	57.9	56.1	57.1	53.2	51.1	59.4	48.8
Last layer	55.0	56.1	56.5	59.4	51.0	49.9	51.7	51.2	55.6	49.6
Proxy-guided fusion (label-free layer choice)
Top-3 avg	57.1-1.2	56.7-1.1	58.5-0.9	60.6+0.7	60.1+3.4	58.2+0.0	54.8-0.6	56.8+0.7	59.0-1.5	49.6+0.0
Top-3 concat	57.0-1.3	56.8-1.0	60.4+1.0	60.8+0.9	60.4+3.7	58.9+0.7	52.7-2.7	55.1-1.0	60.3-0.2	49.6+0.0
Top-5 avg	57.1-1.2	56.7-1.1	58.7-0.7	61.6+1.7	59.1+2.4	58.1-0.1	54.9-0.5	—	61.2+0.7	49.9+0.3
Top-5 concat	58.1-0.1	55.7-2.1	59.6+0.2	59.2-0.7	58.5+1.8	58.0-0.2	53.4-2.0	—	59.8-0.7	49.1-0.5
All-layer fusion (non-trainable)
All-layer avg	57.3-1.0	56.7-1.1	55.8-3.6	59.5-0.4	59.3+2.6	56.2-2.0	52.1-3.3	53.5-2.6	59.3-1.2	50.3+0.7
All-layer concat	56.8-1.5	56.1-1.7	57.0-2.4	58.9-1.0	58.1+1.4	54.6-3.6	52.6-2.8	53.4-2.7	58.0-2.5	50.1+0.5
Trainable fusion
Weighted sum	56.6-1.7	56.7-1.1	56.7-2.7	57.8-2.1	57.8+1.1	54.4-3.8	54.0-1.4	53.6-2.5	60.0-0.5	50.8+1.2
HConv	57.7-0.6	55.8-2.0	56.1-3.3	57.3-2.6	58.9+2.2	57.7-0.5	54.4-1.0	54.5-1.6	59.1-1.4	49.2-0.4
Attentive	57.4-0.9	55.6-2.1	58.6-0.8	59.3-0.6	58.8+2.1	54.3-3.9	50.5-4.9	50.9-5.2	58.1-2.4	49.7+0.1

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (multifeature)	OMAR-RQ (base)
Reference
Oracle (best single layer)	82.6	79.2	84.3	84.8	82.5	80.5
Middle layer	80.8	77.2	83.9	83.5	78.6	75.4
Last layer	79.9	78.4	83.0	82.5	76.9	76.9
Proxy-guided fusion (label-free layer choice)
Top-3 avg	82.1-0.5	—	81.8-2.5	84.7-0.1	82.8+0.3	80.6+0.1
All-layer fusion (non-trainable)
All-layer avg	81.9-0.7	82.0+2.8	83.6-0.7	83.9-0.9	82.2-0.3	79.8-0.7
Trainable fusion
HConv	81.8-0.8	—	84.9+0.6	86.6+1.8	81.7-0.8	80.0-0.5
Attentive	82.3-0.3	—	84.2-0.1	85.9+1.1	82.6+0.1	78.7-1.8

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (multifeature)	OMAR-RQ (base)
Reference
Oracle (best single layer)	62.2	64.3	75.6	78.3	75.4	64.0
Middle layer	54.8	61.3	74.9	75.8	64.0	46.9
Last layer	60.4	54.4	71.9	74.9	69.4	54.1
Proxy-guided fusion (label-free layer choice)
Top-3 avg	60.9-1.3	—	70.8-4.8	77.4-0.9	74.8-0.7	62.3-1.7
All-layer fusion (non-trainable)
All-layer avg	60.6-1.6	60.4-3.9	75.8+0.2	77.3-1.0	75.1-0.3	58.5-5.5
Trainable fusion
HConv	65.3+3.1	—	77.5+1.9	80.0+1.7	77.7+2.2	64.5+0.5
Attentive	64.8+2.5	—	77.6+2.0	79.7+1.4	78.0+2.6	62.9-1.1

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (multifeature)	OMAR-RQ (base)
Reference
Oracle (best single layer)	58.9	60.5	71.5	74.4	70.8	60.3
Middle layer	51.8	57.7	70.8	71.8	59.9	44.4
Last layer	57.1	51.1	67.9	70.7	65.1	51.3
Proxy-guided fusion (label-free layer choice)
Top-3 avg	57.6-1.3	—	67.6-3.9	73.8-0.6	70.6-0.2	58.7-1.6
All-layer fusion (non-trainable)
All-layer avg	57.3-1.6	56.7-3.8	72.1+0.6	73.6-0.8	70.5-0.3	55.2-5.0
Trainable fusion
HConv	62.4+3.5	—	74.3+2.8	76.7+2.3	73.5+2.7	61.0+0.7
Attentive	61.5+2.6	—	73.8+2.3	76.3+1.9	73.3+2.5	59.1-1.2

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B	Myna-Base
Reference
Oracle (best single layer)	64.6	64.2	65.1	65.1	65.5	64.7	62.9	62.9	65.0	57.8
Middle layer	64.6	63.8	64.5	64.2	64.5	64.4	62.3	61.2	64.7	57.4
Last layer	62.1	63.4	64.0	64.3	63.0	60.9	58.9	59.6	63.6	57.8
Proxy-guided fusion (label-free layer choice)
Top-3 avg	64.8+0.2	64.2+0.0	65.0-0.1	64.9-0.2	64.6-0.9	64.7-0.0	63.2+0.3	63.1+0.2	65.5+0.5	57.2-0.5
Top-3 concat	64.8+0.2	63.6-0.6	64.6-0.5	64.4-0.7	63.9-1.6	64.8+0.1	62.9-0.0	62.0-0.9	65.3+0.3	57.1-0.7
Top-5 avg	64.9+0.3	64.2+0.0	65.1+0.0	65.1+0.0	64.6-0.9	64.5-0.2	28.7-34.2	62.6-0.2	65.6+0.6	57.4-0.4
Top-5 concat	64.6+0.0	63.4-0.8	64.7-0.4	64.5-0.6	64.0-1.5	64.6-0.1	41.6-21.2	61.5-1.4	64.8-0.2	56.7-1.0
All-layer fusion (non-trainable)
All-layer avg	64.7+0.1	63.8-0.4	64.8-0.3	65.3+0.2	64.5-1.0	63.5-1.2	62.1-0.8	61.7-1.2	65.1+0.0	57.0-0.8
All-layer concat	64.2-0.4	60.0-4.2	63.9-1.2	64.6-0.5	63.2-2.3	63.2-1.5	60.1-2.8	60.5-2.3	63.4-1.7	55.8-1.9
Trainable fusion
Weighted sum	64.6-0.0	64.7+0.5	65.3+0.2	65.8+0.7	64.6-0.9	63.1-1.6	61.6-1.3	62.5-0.3	65.6+0.5	57.9+0.1
HConv	64.5-0.1	63.4-0.9	64.0-1.1	64.8-0.3	63.6-1.9	62.4-2.4	60.9-2.0	61.1-1.7	64.3-0.8	56.6-1.1
Attentive	65.0+0.4	64.0-0.2	65.4+0.3	65.9+0.8	64.4-1.1	63.6-1.1	62.5-0.4	62.8-0.1	65.1+0.0	56.4-1.3

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B	Myna-Base
Reference
Oracle (best single layer)	50.5	50.1	51.0	51.0	50.6	50.1	47.4	48.0	50.3	44.2
Middle layer	50.5	49.5	50.7	50.6	50.2	49.9	45.0	44.5	50.0	43.8
Last layer	47.1	48.7	49.3	50.2	47.6	45.0	43.0	43.9	48.0	44.2
Proxy-guided fusion (label-free layer choice)
Top-3 avg	50.5+0.0	50.4+0.3	51.4+0.4	52.4+1.4	51.0+0.3	50.4+0.2	48.1+0.7	48.8+0.8	51.3+1.0	43.5-0.7
Top-3 concat	50.3-0.2	49.9-0.2	50.5-0.4	51.4+0.4	50.0-0.6	50.0-0.2	47.5+0.1	47.8-0.2	50.4+0.1	44.0-0.3
Top-5 avg	50.4-0.1	50.0-0.1	51.4+0.4	52.3+1.3	50.9+0.2	50.4+0.2	48.2+0.8	48.9+0.9	51.2+0.9	43.8-0.5
Top-5 concat	50.3-0.2	49.6-0.5	50.1-0.8	50.9-0.1	50.1-0.5	49.9-0.3	47.0-0.4	47.6-0.4	49.8-0.5	44.0-0.2
All-layer fusion (non-trainable)
All-layer avg	50.1-0.4	50.0-0.1	50.3-0.7	51.6+0.6	50.7+0.0	48.8-1.4	46.0-1.4	46.6-1.4	50.3+0.0	43.6-0.6
All-layer concat	49.4-1.1	47.1-3.0	49.8-1.1	50.8-0.2	49.1-1.5	48.4-1.8	44.7-2.7	43.8-4.2	48.1-2.2	43.6-0.7
Trainable fusion
Weighted sum	50.1-0.4	50.0-0.1	50.8-0.2	51.6+0.6	50.6-0.1	49.1-1.1	46.9-0.5	48.3+0.3	51.3+1.0	44.0-0.2
HConv	49.6-0.9	49.1-1.0	50.0-1.0	51.1+0.1	49.1-1.5	48.8-1.3	46.3-1.1	46.7-1.3	49.0-1.3	43.8-0.5
Attentive	49.8-0.7	49.6-0.5	50.5-0.4	51.3+0.3	49.9-0.7	48.7-1.4	47.0-0.4	47.5-0.5	50.1-0.2	43.8-0.4

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B
Reference
Oracle (best single layer)	46.5	46.8	46.9	47.8	46.0	47.2	42.9	43.7	46.2
Middle layer	46.5	45.9	46.8	47.3	45.4	46.9	41.2	40.0	45.4
Last layer	43.9	45.0	45.6	47.3	43.0	42.7	40.1	40.7	44.4
Proxy-guided fusion (label-free layer choice)
Top-3 avg	46.4-0.1	46.3-0.5	47.1+0.3	47.6-0.1	45.9-0.2	46.4-0.8	43.9+1.0	44.8+1.1	46.8+0.5
Top-3 concat	46.4-0.1	46.5-0.3	46.3-0.5	48.2+0.4	45.8-0.2	46.5-0.7	43.7+0.8	43.2-0.5	46.8+0.6
Top-5 avg	46.3-0.2	45.9-0.9	47.2+0.4	47.5-0.3	45.9-0.2	45.5-1.7	43.8+0.9	44.6+0.9	46.7+0.5
Top-5 concat	46.5-0.0	44.6-2.2	46.5-0.3	47.8-0.0	45.7-0.4	45.5-1.7	43.1+0.2	43.2-0.5	46.4+0.2
All-layer fusion (non-trainable)
All-layer avg	45.5-1.0	46.4-0.3	45.2-1.7	47.6-0.2	46.3+0.3	45.6-1.6	41.9-1.0	42.4-1.3	46.0-0.2
All-layer concat	45.7-0.8	43.2-3.6	46.1-0.8	46.9-0.9	45.6-0.4	45.1-2.1	40.4-2.5	40.8-2.9	44.2-2.0
Trainable fusion
Weighted sum	45.6-0.9	46.2-0.5	46.2-0.7	47.2-0.6	45.4-0.7	45.4-1.8	41.4-1.5	42.9-0.8	46.7+0.5
HConv	45.0-1.5	44.1-2.6	44.9-1.9	46.4-1.4	45.7-0.3	45.3-1.9	42.3-0.6	43.5-0.2	45.3-0.9
Attentive	46.0-0.5	45.7-1.0	45.0-1.8	48.0+0.2	45.6-0.4	44.8-2.4	41.9-1.0	43.3-0.4	43.6-2.6

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B	Myna-Base
Reference
Oracle (best single layer)	43.8	43.5	44.9	44.6	43.9	45.0	41.5	42.4	44.3	39.3
Middle layer	43.5	43.1	44.5	44.2	43.5	45.0	39.7	39.4	43.5	39.0
Last layer	41.5	42.6	43.1	43.9	41.1	39.5	37.0	37.9	41.6	39.3
Proxy-guided fusion (label-free layer choice)
Top-3 avg	43.5-0.2	42.9-0.5	44.6-0.3	45.5+0.9	44.4+0.5	45.0-0.0	42.5+0.9	43.0+0.6	44.8+0.5	38.9-0.4
Top-3 concat	43.7-0.0	43.4-0.1	44.3-0.6	44.7+0.1	44.2+0.3	44.5-0.6	42.3+0.8	42.4-0.0	44.1-0.2	39.0-0.3
Top-5 avg	43.7-0.1	43.5+0.0	44.7-0.2	44.5-0.1	44.3+0.4	45.0-0.0	42.6+1.0	43.2+0.8	44.9+0.6	39.1-0.2
Top-5 concat	43.8+0.0	43.7+0.3	43.8-1.1	44.7+0.1	43.9+0.0	44.7-0.3	42.1+0.6	42.3-0.1	43.8-0.5	39.1-0.2
All-layer fusion (non-trainable)
All-layer avg	43.2-0.6	43.6+0.1	43.7-1.2	44.9+0.2	44.1+0.2	43.4-1.7	40.1-1.4	40.6-1.8	43.8-0.5	39.0-0.3
All-layer concat	43.1-0.7	41.0-2.4	43.5-1.4	43.8-0.8	43.3-0.6	42.4-2.7	38.0-3.6	37.6-4.8	40.4-3.9	39.0-0.3
Trainable fusion
Weighted sum	42.7-1.1	43.5+0.1	43.4-1.5	43.9-0.7	43.5-0.4	42.0-3.1	39.1-2.5	41.3-1.1	44.2-0.1	38.7-0.6
HConv	43.1-0.7	42.6-0.9	43.3-1.6	44.3-0.3	42.7-1.2	44.0-1.0	40.8-0.8	40.7-1.7	41.8-2.5	38.5-0.8
Attentive	42.2-1.6	42.4-1.1	44.0-0.9	44.2-0.4	41.8-2.1	41.8-3.3	39.5-2.1	40.0-2.4	42.1-2.2	38.7-0.6

Method	MERT-v1-95M	MERT-v1-330M	MusicFM (MSD)	MuQ (iter)	OMAR-RQ (base)	MusicGen-S	MusicGen-M	MusicGen-L	YuE-s1-0.5B
Reference
Oracle (best single layer)	54.2	54.1	54.4	54.8	54.0	54.2	51.1	51.4	53.5
Middle layer	54.2	53.6	54.3	54.1	53.9	54.0	48.9	47.9	53.3
Last layer	51.4	52.5	53.4	54.4	52.1	49.3	46.8	47.9	52.0
Proxy-guided fusion (label-free layer choice)
Top-3 avg	53.9-0.3	54.2+0.1	54.0-0.3	55.5+0.7	54.1+0.1	53.9-0.3	51.5+0.4	52.3+0.9	54.7+1.3
Top-3 concat	53.6-0.6	53.8-0.3	54.6+0.2	54.8+0.0	53.5-0.5	54.0-0.2	50.6-0.5	51.9+0.5	53.5+0.1
Top-5 avg	53.8-0.4	53.5-0.6	54.3-0.0	55.4+0.6	54.1+0.1	53.9-0.3	51.6+0.5	52.5+1.1	54.6+1.2
Top-5 concat	53.6-0.7	53.1-1.0	54.1-0.2	54.8-0.0	53.4-0.6	53.9-0.3	50.5-0.6	51.5+0.1	52.5-1.0
All-layer fusion (non-trainable)
All-layer avg	53.2-1.0	54.2+0.1	53.4-0.9	55.1+0.3	54.0-0.0	52.1-2.1	49.8-1.4	50.4-1.0	54.2+0.7
All-layer concat	53.0-1.3	51.6-2.5	54.1-0.2	53.7-1.1	53.1-0.9	52.3-1.9	47.2-3.9	47.9-3.5	50.9-2.6
Trainable fusion
Weighted sum	53.2-1.0	54.1-0.0	53.8-0.5	54.7-0.1	54.2+0.2	53.5-0.7	51.2+0.1	52.2+0.8	54.7+1.2
HConv	53.5-0.8	53.7-0.4	53.4-0.9	54.5-0.3	52.5-1.5	51.3-2.9	50.0-1.1	50.3-1.1	52.6-0.9
Attentive	53.8-0.5	53.8-0.3	54.3-0.1	55.5+0.7	54.2+0.2	53.2-1.0	50.8-0.3	51.2-0.2	53.8+0.3

Method	Key	Pitch	Instrument	Genre	Beat	Downbeat	Emotion	Structure	Chords	Chords root	Chords thirds	MTT	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50
Reference
Oracle (best single layer)	63.4	92.4	70.3	78.6	87.7	47.5	66.3	58.3	82.6	62.2	58.9	64.6	50.5	46.5	43.8	54.2
Middle layer	61.1	90.3	70.0	77.2	87.6	47.1	65.2	57.8	80.8	54.8	51.8	64.6	50.5	46.5	43.5	54.2
Last layer	61.0	90.8	67.7	65.9	86.0	44.5	49.7	55.0	79.9	60.4	57.1	62.1	47.1	43.9	41.5	51.4
Proxy-guided fusion (label-free layer choice)
Top-3 avg	61.8-1.6	—	69.6-0.7	77.6-1.0	87.4-0.3	47.3-0.2	63.3-3.0	57.1-1.2	82.1-0.5	60.9-1.3	57.6-1.3	64.8+0.2	50.5+0.0	46.4-0.1	43.5-0.2	53.9-0.3
Top-3 concat	62.8-0.6	—	70.5+0.2	67.6-11.0	87.8+0.1	47.4-0.1	63.2-3.1	57.0-1.3	—	—	—	64.8+0.2	50.3-0.2	46.4-0.1	43.7-0.0	53.6-0.6
Top-5 avg	62.1-1.3	—	69.0-1.3	77.9-0.7	87.3-0.4	47.3-0.2	62.7-3.6	57.1-1.2	—	—	—	64.9+0.3	50.4-0.1	46.3-0.2	43.7-0.1	53.8-0.4
Top-5 concat	60.4-3.0	—	70.7+0.4	76.9-1.7	87.0-0.7	46.6-0.9	64.3-2.0	58.1-0.1	—	—	—	64.6+0.0	50.3-0.2	46.5-0.0	43.8+0.0	53.6-0.7
All-layer fusion (non-trainable)
All-layer avg	64.8+1.4	92.4-0.0	69.7-0.6	75.5-3.1	86.1-1.6	46.4-1.1	59.7-6.6	57.3-1.0	81.9-0.7	60.6-1.6	57.3-1.6	64.7+0.1	50.1-0.4	45.5-1.0	43.2-0.6	53.2-1.0
All-layer concat	59.7-3.7	—	70.6+0.3	72.8-5.8	87.2-0.5	47.2-0.3	62.5-3.8	56.8-1.5	—	—	—	64.2-0.4	49.4-1.1	45.7-0.8	43.1-0.7	53.0-1.3
Trainable fusion
Weighted sum	62.0-1.4	—	69.1-1.2	66.5-12.0	86.9-0.8	46.8-0.7	55.7-10.6	56.6-1.7	—	—	—	64.6-0.0	50.1-0.4	45.6-0.9	42.7-1.1	53.2-1.0
HConv	60.6-2.8	90.4-2.0	71.7+1.4	67.6-11.0	87.5-0.2	47.7+0.2	57.1-9.2	57.7-0.6	81.8-0.8	65.3+3.1	62.4+3.5	64.5-0.1	49.6-0.9	45.0-1.5	43.1-0.7	53.5-0.8
Attentive	57.2-6.2	92.8+0.3	68.1-2.2	69.7-8.9	88.0+0.3	47.1-0.4	59.1-7.2	57.4-0.9	82.3-0.3	64.8+2.5	61.5+2.6	65.0+0.4	49.8-0.7	46.0-0.5	42.2-1.6	53.8-0.5

Method	Key	Pitch	Instrument	Genre	Beat	Downbeat	Emotion	Structure	Chords	Chords root	Chords thirds	MTT	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50
Reference
Oracle (best single layer)	65.2	92.2	71.4	78.3	88.0	47.8	66.4	57.8	79.2	64.3	60.5	64.2	50.1	46.8	43.5	54.1
Middle layer	64.3	89.9	69.4	74.5	87.7	46.9	63.7	56.1	77.2	61.3	57.7	63.8	49.5	45.9	43.1	53.6
Last layer	58.4	90.8	67.7	72.4	87.0	46.1	63.0	56.1	78.4	54.4	51.1	63.4	48.7	45.0	42.6	52.5
Proxy-guided fusion (label-free layer choice)
Top-3 avg	63.6-1.6	—	68.9-2.5	75.2-3.1	87.3-0.7	46.9-0.9	63.7-2.7	56.7-1.1	—	—	—	64.2+0.0	50.4+0.3	46.3-0.5	42.9-0.5	54.2+0.1
Top-3 concat	63.7-1.5	—	69.4-2.0	74.5-3.8	87.6-0.4	47.7-0.1	62.3-4.1	56.8-1.0	—	—	—	63.6-0.6	49.9-0.2	46.5-0.3	43.4-0.1	53.8-0.3
Top-5 avg	65.2+0.0	—	68.9-2.5	72.8-5.5	87.1-0.9	46.0-1.8	64.5-1.9	56.7-1.1	—	—	—	64.2+0.0	50.0-0.1	45.9-0.9	43.5+0.0	53.5-0.6
Top-5 concat	61.8-3.5	—	70.1-1.3	73.8-4.5	87.4-0.6	46.5-1.3	61.1-5.3	55.7-2.1	—	—	—	63.4-0.8	49.6-0.5	44.6-2.2	43.7+0.3	53.1-1.0
All-layer fusion (non-trainable)
All-layer avg	65.4+0.2	90.9-1.3	69.2-2.2	70.7-7.6	87.1-0.9	47.0-0.8	61.2-5.2	56.7-1.1	82.0+2.8	60.4-3.9	56.7-3.8	63.8-0.4	50.0-0.1	46.4-0.3	43.6+0.1	54.2+0.1
All-layer concat	26.7-38.5	—	66.5-4.9	67.2-11.1	87.2-0.8	47.3-0.5	61.2-5.2	56.1-1.7	—	—	—	60.0-4.2	47.1-3.0	43.2-3.6	41.0-2.4	51.6-2.5
Trainable fusion
Weighted sum	62.9-2.3	—	70.0-1.4	68.3-10.0	87.3-0.7	46.7-1.1	62.3-4.1	56.7-1.1	—	—	—	64.7+0.5	50.0-0.1	46.2-0.5	43.5+0.1	54.1-0.0
HConv	60.1-5.1	90.2-2.0	69.2-2.2	49.0-29.3	87.3-0.7	46.8-1.0	61.0-5.5	55.8-2.0	—	—	—	63.4-0.9	49.1-1.0	44.1-2.6	42.6-0.9	53.7-0.4
Attentive	61.5-3.7	91.5-0.7	64.6-6.8	64.1-14.2	88.6+0.6	47.6-0.2	58.1-8.3	55.6-2.1	—	—	—	64.0-0.2	49.6-0.5	45.7-1.0	42.4-1.1	53.8-0.3

Method	Key	Pitch	Instrument	Genre	Beat	Downbeat	Emotion	Structure	Chords	Chords root	Chords thirds	MTT	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50
Reference
Oracle (best single layer)	64.6	90.5	75.5	84.8	75.7	45.9	68.4	59.4	84.3	75.6	71.5	65.1	51.0	46.9	44.9	54.4
Middle layer	64.3	89.0	74.1	83.1	73.9	45.0	67.4	57.1	83.9	74.9	70.8	64.5	50.7	46.8	44.5	54.3
Last layer	62.2	86.9	72.0	69.7	73.5	44.8	64.2	56.5	83.0	71.9	67.9	64.0	49.3	45.6	43.1	53.4
Proxy-guided fusion (label-free layer choice)
Top-3 avg	59.5-5.1	—	73.6-1.9	81.4-3.4	75.5-0.2	44.4-1.5	64.6-3.8	58.5-0.9	81.8-2.5	70.8-4.8	67.6-3.9	65.0-0.1	51.4+0.4	47.1+0.3	44.6-0.3	54.0-0.3
Top-3 concat	58.5-6.1	—	75.1-0.4	83.1-1.7	76.2+0.5	43.8-2.1	66.2-2.2	60.4+1.0	—	—	—	64.6-0.5	50.5-0.4	46.3-0.5	44.3-0.6	54.6+0.2
Top-5 avg	60.8-3.8	—	74.6-0.9	80.7-4.1	75.3-0.4	44.4-1.5	66.3-2.1	58.7-0.7	—	—	—	65.1+0.0	51.4+0.4	47.2+0.4	44.7-0.2	54.3-0.0
Top-5 concat	61.3-3.3	—	74.5-1.0	82.4-2.4	76.5+0.8	44.0-1.9	66.5-1.9	59.6+0.2	—	—	—	64.7-0.4	50.1-0.8	46.5-0.3	43.8-1.1	54.1-0.2
All-layer fusion (non-trainable)
All-layer avg	63.5-1.1	90.8+0.2	75.5+0.0	71.4-13.4	74.2-1.5	46.0+0.1	62.1-6.3	55.8-3.6	83.6-0.7	75.8+0.2	72.1+0.6	64.8-0.3	50.3-0.7	45.2-1.7	43.7-1.2	53.4-0.9
All-layer concat	61.0-3.6	—	76.2+0.7	71.0-13.8	79.2+3.5	43.2-2.7	65.6-2.8	57.0-2.4	—	—	—	63.9-1.2	49.8-1.1	46.1-0.8	43.5-1.4	54.1-0.2
Trainable fusion
Weighted sum	62.2-2.3	—	75.2-0.3	63.1-21.7	74.8-0.9	44.0-1.9	55.7-12.7	56.7-2.7	—	—	—	65.3+0.2	50.8-0.2	46.2-0.7	43.4-1.5	53.8-0.5
HConv	61.8-2.8	89.7-0.8	74.2-1.3	64.1-20.7	83.8+8.1	42.3-3.6	59.9-8.5	56.1-3.3	84.9+0.6	77.5+1.9	74.3+2.8	64.0-1.1	50.0-1.0	44.9-1.9	43.3-1.6	53.4-0.9
Attentive	59.8-4.8	91.8+1.2	73.9-1.6	79.0-5.8	80.6+4.9	45.8-0.1	58.8-9.6	58.6-0.8	84.2-0.1	77.6+2.0	73.8+2.3	65.4+0.3	50.5-0.4	45.0-1.8	44.0-0.9	54.3-0.1

Method	Key	Pitch	Instrument	Genre	Beat	Downbeat	Emotion	Structure	Chords	Chords root	Chords thirds	MTT	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50
Reference
Oracle (best single layer)	64.5	92.5	79.1	85.2	75.8	46.5	67.1	59.9	84.8	78.3	74.4	65.1	51.0	47.8	44.6	54.8
Middle layer	61.9	90.1	78.3	84.1	75.8	42.6	67.0	57.9	83.5	75.8	71.8	64.2	50.6	47.3	44.2	54.1
Last layer	58.0	87.6	75.2	80.7	74.8	46.5	64.0	59.4	82.5	74.9	70.7	64.3	50.2	47.3	43.9	54.4
Proxy-guided fusion (label-free layer choice)
Top-3 avg	64.4-0.1	—	78.8-0.3	84.8-0.4	74.3-1.5	46.4-0.1	67.2+0.1	60.6+0.7	84.7-0.1	77.4-0.9	73.8-0.6	64.9-0.2	52.4+1.4	47.6-0.1	45.5+0.9	55.5+0.7
Top-3 concat	63.3-1.2	—	79.7+0.6	82.1-3.1	79.9+4.1	42.6-3.9	70.5+3.4	60.8+0.9	—	—	—	64.4-0.7	51.4+0.4	48.2+0.4	44.7+0.1	54.8+0.0
Top-5 avg	63.7-0.8	—	78.4-0.7	84.8-0.4	75.0-0.8	45.6-0.9	66.4-0.7	61.6+1.7	—	—	—	65.1+0.0	52.3+1.3	47.5-0.3	44.5-0.1	55.4+0.6
Top-5 concat	61.2-3.3	—	78.5-0.6	84.1-1.1	79.1+3.3	42.1-4.4	67.5+0.4	59.2-0.7	—	—	—	64.5-0.6	50.9-0.1	47.8-0.0	44.7+0.1	54.8-0.0
All-layer fusion (non-trainable)
All-layer avg	64.5+0.0	91.9-0.6	78.7-0.4	84.1-1.1	76.9+1.1	44.1-2.4	66.1-1.0	59.5-0.4	83.9-0.9	77.3-1.0	73.6-0.8	65.3+0.2	51.6+0.6	47.6-0.2	44.9+0.2	55.1+0.3
All-layer concat	63.3-1.2	—	80.8+1.7	79.0-6.2	82.4+6.6	39.5-7.0	69.0+2.0	58.9-1.0	—	—	—	64.6-0.5	50.8-0.2	46.9-0.9	43.8-0.8	53.7-1.1
Trainable fusion
Weighted sum	63.1-1.4	—	81.2+2.2	70.3-14.9	78.0+2.2	42.1-4.4	65.0-2.1	57.8-2.1	—	—	—	65.8+0.7	51.6+0.6	47.2-0.6	43.9-0.7	54.7-0.1
HConv	63.0-1.5	91.5-1.0	78.4-0.7	80.7-4.5	84.7+8.9	43.6-2.9	68.9+1.8	57.3-2.6	86.6+1.8	80.0+1.7	76.7+2.3	64.8-0.3	51.1+0.1	46.4-1.4	44.3-0.3	54.5-0.3
Attentive	61.7-2.8	92.8+0.3	77.2-1.9	84.5-0.7	80.2+4.4	45.0-1.5	57.3-9.8	59.3-0.6	85.9+1.1	79.7+1.4	76.3+1.9	65.9+0.8	51.3+0.3	48.0+0.2	44.2-0.4	55.5+0.7

Method	Chords	Chords root	Chords thirds
Reference
Oracle (best single layer)	82.5	75.4	70.8
Middle layer	78.6	64.0	59.9
Last layer	76.9	69.4	65.1
Proxy-guided fusion (label-free layer choice)
Top-3 avg	82.8+0.3	74.8-0.7	70.6-0.2
All-layer fusion (non-trainable)
All-layer avg	82.2-0.3	75.1-0.3	70.5-0.3
Trainable fusion
HConv	81.7-0.8	77.7+2.2	73.5+2.7
Attentive	82.6+0.1	78.0+2.6	73.3+2.5

Method	Key	Pitch	Instrument	Genre	Beat	Downbeat	Emotion	Structure	Chords	Chords root	Chords thirds	MTT	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50
Reference
Oracle (best single layer)	62.2	90.9	74.8	81.0	73.7	43.3	66.0	56.7	80.5	64.0	60.3	65.5	50.6	46.0	43.9	54.0
Middle layer	55.6	88.1	73.5	74.8	66.0	43.3	64.2	56.1	75.4	46.9	44.4	64.5	50.2	45.4	43.5	53.9
Last layer	51.1	84.4	68.3	59.7	55.5	41.5	59.3	51.0	76.9	54.1	51.3	63.0	47.6	43.0	41.1	52.1
Proxy-guided fusion (label-free layer choice)
Top-3 avg	66.2+4.0	—	72.9-1.9	80.7-0.3	74.5+0.8	46.0+2.7	67.7+1.7	60.1+3.4	80.6+0.1	62.3-1.7	58.7-1.6	64.6-0.9	51.0+0.3	45.9-0.2	44.4+0.5	54.1+0.1
Top-3 concat	65.3+3.1	—	74.2-0.6	80.7-0.3	76.8+3.1	44.8+1.5	66.8+0.8	60.4+3.7	—	—	—	63.9-1.6	50.0-0.6	45.8-0.2	44.2+0.3	53.5-0.5
Top-5 avg	66.6+4.4	—	72.3-2.5	76.9-4.1	75.0+1.3	45.3+2.0	68.0+2.0	59.1+2.4	—	—	—	64.6-0.9	50.9+0.2	45.9-0.2	44.3+0.4	54.1+0.1
Top-5 concat	60.2-2.0	—	73.1-1.7	79.0-2.0	76.7+3.0	44.2+0.9	67.6+1.6	58.5+1.8	—	—	—	64.0-1.5	50.1-0.5	45.7-0.4	43.9+0.0	53.4-0.6
All-layer fusion (non-trainable)
All-layer avg	63.3+1.1	94.1+3.2	74.1-0.7	76.2-4.8	74.3+0.6	46.4+3.1	67.5+1.5	59.3+2.6	79.8-0.7	58.5-5.5	55.2-5.0	64.5-1.0	50.7+0.0	46.3+0.3	44.1+0.2	54.0-0.0
All-layer concat	61.7-0.5	—	75.1+0.3	70.0-11.0	81.6+7.9	40.3-3.0	65.6-0.4	58.1+1.4	—	—	—	63.2-2.3	49.1-1.5	45.6-0.4	43.3-0.6	53.1-0.9
Trainable fusion
Weighted sum	64.8+2.6	—	74.7-0.1	67.9-13.1	79.4+5.7	41.2-2.1	61.4-4.6	57.8+1.1	—	—	—	64.6-0.9	50.6-0.1	45.4-0.7	43.5-0.4	54.2+0.2
HConv	63.8+1.6	92.2+1.3	75.2+0.4	67.6-13.4	84.5+10.8	43.8+0.5	59.9-6.1	58.9+2.2	80.0-0.5	64.5+0.5	61.0+0.7	63.6-1.9	49.1-1.5	45.7-0.3	42.7-1.2	52.5-1.5
Attentive	60.6-1.6	94.4+3.5	70.0-4.8	77.2-3.8	79.6+5.9	46.5+3.2	58.5-7.5	58.8+2.1	78.7-1.8	62.9-1.1	59.1-1.2	64.4-1.1	49.9-0.7	45.6-0.4	41.8-2.1	54.2+0.2

Method	Key	Genre	Beat	Downbeat	Emotion	Structure	MTT	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50
Reference
Oracle (best single layer)	68.6	74.8	85.7	47.7	64.5	58.2	64.7	50.1	47.2	45.0	54.2
Middle layer	66.3	70.7	85.5	47.5	63.0	57.1	64.4	49.9	46.9	45.0	54.0
Last layer	63.2	52.4	82.6	40.6	45.1	49.9	60.9	45.0	42.7	39.5	49.3
Proxy-guided fusion (label-free layer choice)
Top-3 avg	67.8-0.8	74.8+0.0	85.3-0.4	47.3-0.4	64.0-0.5	58.2+0.0	64.7-0.0	50.4+0.2	46.4-0.8	45.0-0.0	53.9-0.3
Top-3 concat	66.9-1.7	72.4-2.4	85.9+0.2	47.3-0.4	55.2-9.3	58.9+0.7	64.8+0.1	50.0-0.2	46.5-0.7	44.5-0.6	54.0-0.2
Top-5 avg	67.7-0.9	73.8-1.0	85.2-0.5	47.1-0.6	64.4-0.1	58.1-0.1	64.5-0.2	50.4+0.2	45.5-1.7	45.0-0.0	53.9-0.3
Top-5 concat	66.4-2.2	73.1-1.7	86.1+0.4	47.6-0.1	63.6-0.9	58.0-0.2	64.6-0.1	49.9-0.3	45.5-1.7	44.7-0.3	53.9-0.3
All-layer fusion (non-trainable)
All-layer avg	68.0-0.6	56.9-17.9	84.8-0.9	46.3-1.4	55.4-9.1	56.2-2.0	63.5-1.2	48.8-1.4	45.6-1.6	43.4-1.7	52.1-2.1
All-layer concat	66.0-2.6	53.8-21.0	85.6-0.1	46.9-0.8	61.8-2.7	54.6-3.6	63.2-1.5	48.4-1.8	45.1-2.1	42.4-2.7	52.3-1.9
Trainable fusion
Weighted sum	66.2-2.4	53.5-21.3	85.7+0.0	47.4-0.3	52.8-11.7	54.4-3.8	63.1-1.6	49.1-1.1	45.4-1.8	42.0-3.1	53.5-0.7
HConv	62.9-5.7	58.6-16.2	85.5-0.2	47.0-0.7	60.2-4.3	57.7-0.5	62.4-2.4	48.8-1.3	45.3-1.9	44.0-1.0	51.3-2.9
Attentive	62.0-6.6	59.3-15.5	86.0+0.3	47.1-0.6	50.3-14.2	54.3-3.9	63.6-1.1	48.7-1.4	44.8-2.4	41.8-3.3	53.2-1.0

Method	Key	Pitch	Instrument	Genre	Beat	Downbeat	Emotion	Structure	MTT	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50
Reference
Oracle (best single layer)	65.5	89.8	69.4	60.7	85.0	45.7	53.8	55.4	62.9	47.4	42.9	41.5	51.1
Middle layer	55.2	86.7	67.2	45.9	84.2	44.5	50.2	53.2	62.3	45.0	41.2	39.7	48.9
Last layer	63.4	86.0	64.9	42.4	81.8	40.5	31.2	51.7	58.9	43.0	40.1	37.0	46.8
Proxy-guided fusion (label-free layer choice)
Top-3 avg	65.8+0.3	—	69.1-0.3	60.7-0.0	84.8-0.2	45.8+0.1	49.5-4.3	54.8-0.6	63.2+0.3	48.1+0.7	43.9+1.0	42.5+0.9	51.5+0.4
Top-3 concat	63.5-2.0	—	68.8-0.6	60.3-0.4	84.6-0.4	45.2-0.5	44.5-9.3	52.7-2.7	62.9-0.0	47.5+0.1	43.7+0.8	42.3+0.8	50.6-0.5
Top-5 avg	66.6+1.1	—	69.2-0.2	60.3-0.4	84.5-0.5	45.7-0.0	49.5-4.2	54.9-0.5	28.7-34.2	48.2+0.8	43.8+0.9	42.6+1.0	51.6+0.5
Top-5 concat	63.9-1.6	—	67.7-1.7	62.4+1.7	84.8-0.2	45.6-0.1	50.4-3.4	53.4-2.0	41.6-21.2	47.0-0.4	43.1+0.2	42.1+0.6	50.5-0.6
All-layer fusion (non-trainable)
All-layer avg	63.7-1.8	88.6-1.2	67.7-1.7	40.3-20.4	84.0-1.0	44.5-1.2	44.6-9.2	52.1-3.3	62.1-0.8	46.0-1.4	41.9-1.0	40.1-1.4	49.8-1.4
All-layer concat	62.3-3.2	—	69.4+0.0	40.0-20.7	84.3-0.7	45.0-0.8	51.3-2.5	52.6-2.8	60.1-2.8	44.7-2.7	40.4-2.5	38.0-3.6	47.2-3.9
Trainable fusion
Weighted sum	65.6+0.1	—	69.9+0.5	38.6-22.1	84.5-0.5	44.8-0.9	39.4-14.4	54.0-1.4	61.6-1.3	46.9-0.5	41.4-1.5	39.1-2.5	51.2+0.1
HConv	62.4-3.1	87.3-2.5	67.4-2.0	37.6-23.1	84.4-0.6	45.4-0.3	45.1-8.7	54.4-1.0	60.9-2.0	46.3-1.1	42.3-0.6	40.8-0.8	50.0-1.1
Attentive	58.1-7.4	91.1+1.3	69.2-0.2	44.8-15.9	84.6-0.4	45.4-0.3	48.0-5.8	50.5-4.9	62.5-0.4	47.0-0.4	41.9-1.0	39.5-2.1	50.8-0.3

Method	Key	Pitch	Instrument	Genre	Beat	Downbeat	Emotion	Structure	MTT	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50
Reference
Oracle (best single layer)	67.2	90.9	69.8	63.8	85.3	46.5	56.1	56.1	62.9	48.0	43.7	42.4	51.4
Middle layer	54.1	85.6	68.8	43.4	82.5	43.0	48.3	51.1	61.2	44.5	40.0	39.4	47.9
Last layer	64.9	88.7	66.7	44.1	81.5	42.3	36.6	51.2	59.6	43.9	40.7	37.9	47.9
Proxy-guided fusion (label-free layer choice)
Top-3 avg	—	—	—	60.7-3.1	84.7-0.6	46.1-0.4	54.0-2.1	56.8+0.7	63.1+0.2	48.8+0.8	44.8+1.1	43.0+0.6	52.3+0.9
Top-3 concat	—	—	—	60.0-3.8	85.0-0.3	46.0-0.5	52.1-4.0	55.1-1.0	62.0-0.9	47.8-0.2	43.2-0.5	42.4-0.0	51.9+0.5
Top-5 avg	68.2+1.0	—	68.7-1.1	60.7-3.1	84.4-0.9	45.9-0.6	54.2-1.9	—	62.6-0.2	48.9+0.9	44.6+0.9	43.2+0.8	52.5+1.1
Top-5 concat	66.5-0.7	—	69.0-0.8	56.9-6.9	85.0-0.3	45.7-0.8	55.9-0.2	—	61.5-1.4	47.6-0.4	43.2-0.5	42.3-0.1	51.5+0.1
All-layer fusion (non-trainable)
All-layer avg	65.8-1.5	89.3-1.6	69.9+0.1	40.3-23.5	84.5-0.8	44.9-1.6	50.4-5.7	53.5-2.6	61.7-1.2	46.6-1.4	42.4-1.3	40.6-1.8	50.4-1.0
All-layer concat	62.4-4.8	—	69.4-0.4	45.2-18.6	84.6-0.7	45.1-1.4	51.7-4.4	53.4-2.7	60.5-2.3	43.8-4.2	40.8-2.9	37.6-4.8	47.9-3.5
Trainable fusion
Weighted sum	65.6-1.6	—	69.8+0.0	38.3-25.5	84.7-0.6	46.1-0.4	50.1-6.0	53.6-2.5	62.5-0.3	48.3+0.3	42.9-0.8	41.3-1.1	52.2+0.8
HConv	60.6-6.6	87.6-3.3	67.9-1.9	43.8-20.0	84.4-0.9	45.1-1.4	50.1-6.0	54.5-1.6	61.1-1.7	46.7-1.3	43.5-0.2	40.7-1.7	50.3-1.1
Attentive	60.2-7.0	91.6+0.7	70.5+0.7	42.8-21.0	85.1-0.2	45.4-1.1	35.1-21.0	50.9-5.2	62.8-0.1	47.5-0.5	43.3-0.4	40.0-2.4	51.2-0.2

Method	Key	Genre	Beat	Downbeat	Emotion	Structure	MTT	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50
Reference
Oracle (best single layer)	66.6	81.4	88.2	48.8	66.7	60.5	65.0	50.3	46.2	44.3	53.5
Middle layer	54.3	81.4	88.1	48.7	64.5	59.4	64.7	50.0	45.4	43.5	53.3
Last layer	66.6	66.6	86.5	46.9	60.5	55.6	63.6	48.0	44.4	41.6	52.0
Proxy-guided fusion (label-free layer choice)
Top-3 avg	67.0+0.4	81.7+0.3	87.7-0.5	48.4-0.4	67.9+1.2	59.0-1.5	65.5+0.5	51.3+1.0	46.8+0.5	44.8+0.5	54.7+1.3
Top-3 concat	65.0-1.6	80.3-1.1	87.9-0.3	48.4-0.4	64.5-2.2	60.3-0.2	65.3+0.3	50.4+0.1	46.8+0.6	44.1-0.2	53.5+0.1
Top-5 avg	67.4+0.8	81.7+0.3	87.7-0.5	48.3-0.5	68.4+1.7	61.2+0.7	65.6+0.6	51.2+0.9	46.7+0.5	44.9+0.6	54.6+1.2
Top-5 concat	62.5-4.1	73.8-7.6	87.6-0.6	48.5-0.3	64.5-2.2	59.8-0.7	64.8-0.2	49.8-0.5	46.4+0.2	43.8-0.5	52.5-1.0
All-layer fusion (non-trainable)
All-layer avg	68.0+1.4	74.8-6.6	88.0-0.2	48.5-0.3	66.0-0.7	59.3-1.2	65.1+0.0	50.3+0.0	46.0-0.2	43.8-0.5	54.2+0.7
All-layer concat	61.8-4.8	68.3-13.1	88.1-0.1	48.1-0.7	65.2-1.5	58.0-2.5	63.4-1.7	48.1-2.2	44.2-2.0	40.4-3.9	50.9-2.6
Trainable fusion
Weighted sum	64.4-2.2	71.7-9.7	88.1-0.1	48.3-0.5	66.5-0.2	60.0-0.5	65.6+0.5	51.3+1.0	46.7+0.5	44.2-0.1	54.7+1.2
HConv	61.4-5.2	42.1-39.3	88.0-0.2	48.5-0.3	60.5-6.2	59.1-1.4	64.3-0.8	49.0-1.3	45.3-0.9	41.8-2.5	52.6-0.9
Attentive	61.7-4.9	68.3-13.1	88.4+0.2	48.8+0.0	55.7-11.0	58.1-2.4	65.1+0.0	50.1-0.2	43.6-2.6	42.1-2.2	53.8+0.3

Method	Key	Instrument	Genre	Emotion	Structure	MTT	MTG-Genre	MTG-Mood
Reference
Oracle (best single layer)	47.8	64.5	52.1	43.2	49.6	57.8	44.2	39.3
Middle layer	47.8	63.0	49.0	41.3	48.8	57.4	43.8	39.0
Last layer	46.7	64.5	49.3	43.2	49.6	57.8	44.2	39.3
Proxy-guided fusion (label-free layer choice)
Top-3 avg	45.5-2.3	63.9-0.6	46.5-5.6	37.6-5.6	49.6+0.0	57.2-0.5	43.5-0.7	38.9-0.4
Top-3 concat	45.1-2.7	62.2-2.3	46.5-5.6	41.0-2.2	49.6+0.0	57.1-0.7	44.0-0.3	39.0-0.3
Top-5 avg	47.0-0.8	63.0-1.5	46.9-5.2	41.2-2.0	49.9+0.3	57.4-0.4	43.8-0.5	39.1-0.2
Top-5 concat	46.7-1.1	59.2-5.3	22.4-29.7	39.9-3.3	49.1-0.5	56.7-1.0	44.0-0.2	39.1-0.2
All-layer fusion (non-trainable)
All-layer avg	46.7-1.1	64.6+0.1	44.5-7.6	38.7-4.5	50.3+0.7	57.0-0.8	43.6-0.6	39.0-0.3
All-layer concat	40.6-7.2	54.7-9.8	22.8-29.3	40.2-3.0	50.1+0.5	55.8-1.9	43.6-0.7	39.0-0.3
Trainable fusion
Weighted sum	47.6-0.2	64.8+0.3	46.2-5.9	34.0-9.2	50.8+1.2	57.9+0.1	44.0-0.2	38.7-0.6
HConv	48.0+0.2	65.9+1.4	49.0-3.1	29.8-13.5	49.2-0.4	56.6-1.1	43.8-0.5	38.5-0.8
Attentive	39.4-8.4	58.6-5.9	47.9-4.2	34.7-8.5	49.7+0.1	56.4-1.3	43.8-0.4	38.7-0.6