Metric–Task Correlations

Which label-free metrics track layer quality on which tasks? Each cell: the mean, over models, of the within-model Spearman correlation between a metric’s layer profile and a task’s per-layer performance. Hover a cell for the per-model values.

← Results table · Layer cheat sheet · Layer explorer · Selection & fusion · Paper · results as of 2026-08-04

Models

positive (higher metric → better layers) negative · saturation tracks |ρ| · — insufficient data · same polarity as the paper’s figure

Metric	Chords MIREX	Chords root	Chords thirds	Chords triads	Key	Pitch	Beat	Downbeat	Instrument	Emotion	Genre	Structure	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50	MTT	DimSim
ID (GRIDE k=8)	+0.32	+0.36	+0.34	+0.26	+0.21	+0.03	+0.52	+0.54	+0.49	+0.60	+0.56	+0.54	+0.64	+0.65	+0.61	+0.63	+0.60	+0.55
ID (TwoNN)	+0.23	+0.27	+0.24	+0.16	+0.14	-0.02	+0.50	+0.43	+0.48	+0.54	+0.45	+0.48	+0.42	+0.43	+0.40	+0.39	+0.41	+0.38
Alpha-ReQ	-0.25	-0.14	-0.13	-0.07	+0.12	+0.19	-0.51	-0.37	-0.54	-0.58	-0.54	-0.36	-0.38	-0.36	-0.36	-0.32	-0.40	-0.26
Effective rank	+0.27	+0.25	+0.24	+0.18	-0.09	-0.22	+0.50	+0.43	+0.36	+0.49	+0.48	+0.40	+0.34	+0.31	+0.32	+0.35	+0.39	+0.16
Anisotropy	-0.35	-0.51	-0.49	-0.45	-0.09	-0.01	-0.21	-0.07	-0.32	-0.21	-0.15	-0.12	-0.13	-0.08	-0.08	-0.07	-0.11	-0.31
InfoNCE	-0.17	-0.04	-0.02	+0.05	+0.02	+0.20	-0.73	-0.65	-0.45	-0.61	-0.53	-0.64	-0.61	-0.55	-0.60	-0.63	-0.45	-0.42
LiDAR	-0.05	-0.08	-0.10	-0.16	-0.39	-0.53	+0.48	+0.41	+0.20	+0.43	+0.44	+0.40	+0.46	+0.32	+0.43	+0.47	+0.30	-0.20
Curvature	-0.12	-0.12	-0.12	-0.05	-0.00	+0.27	-0.39	-0.49	-0.19	-0.31	-0.27	-0.34	-0.39	-0.35	-0.33	-0.46	-0.17	-0.12
PTE	+0.24	+0.57	+0.59	+0.61	+0.71	+0.74	-0.12	+0.03	+0.29	+0.14	+0.00	+0.14	+0.14	+0.22	+0.15	+0.10	+0.12	+0.41

Metric	Chords MIREX	Chords root	Chords thirds	Chords triads	Key	Pitch	Beat	Downbeat	Instrument	Emotion	Genre	Structure	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50	MTT	DimSim
ID (GRIDE k=8)	+0.32	+0.36	+0.34	+0.26	+0.22	-0.08	+0.41	+0.44	+0.39	+0.60	+0.58	+0.45	+0.55	+0.58	+0.51	+0.53	+0.51	+0.46
ID (TwoNN)	+0.23	+0.24	+0.21	+0.14	+0.10	-0.21	+0.43	+0.46	+0.30	+0.61	+0.52	+0.47	+0.49	+0.50	+0.45	+0.49	+0.44	+0.34
Alpha-ReQ	-0.24	-0.08	-0.08	-0.02	+0.06	+0.21	-0.53	-0.40	-0.45	-0.58	-0.64	-0.34	-0.46	-0.47	-0.43	-0.44	-0.37	-0.37
Effective rank	+0.26	+0.22	+0.21	+0.15	-0.02	-0.37	+0.32	+0.31	+0.06	+0.30	+0.30	+0.25	+0.15	+0.14	+0.11	+0.24	+0.19	-0.08
Anisotropy	-0.30	-0.44	-0.43	-0.40	-0.21	-0.10	-0.21	-0.10	-0.22	-0.21	-0.26	-0.07	-0.22	-0.19	-0.19	-0.18	-0.17	-0.25
InfoNCE	-0.14	+0.02	+0.04	+0.12	+0.20	+0.58	-0.69	-0.62	-0.28	-0.61	-0.66	-0.67	-0.63	-0.51	-0.64	-0.72	-0.47	-0.27
LiDAR	-0.05	-0.08	-0.10	-0.16	-0.16	-0.52	+0.87	+0.74	+0.33	+0.65	+0.70	+0.71	+0.78	+0.53	+0.73	+0.83	+0.42	+0.04
Curvature	-0.11	-0.12	-0.11	-0.03	+0.01	+0.33	-0.54	-0.67	-0.31	-0.56	-0.56	-0.65	-0.68	-0.57	-0.59	-0.70	-0.46	-0.28
PTE	+0.23	+0.58	+0.59	+0.62	+0.85	+0.60	-0.52	-0.48	+0.11	-0.36	-0.30	-0.26	-0.30	-0.11	-0.29	-0.38	-0.07	+0.12

Metric	Chords MIREX	Chords root	Chords thirds	Chords triads	Key	Pitch	Beat	Downbeat	Instrument	Emotion	Genre	Structure	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50	MTT	DimSim
ID (GRIDE k=8)	—	—	—	—	+0.67	+0.53	+0.83	+0.85	+0.73	+0.80	+0.81	+0.67	+0.78	+0.74	+0.77	+0.78	+0.75	+0.70
ID (TwoNN)	—	—	—	—	+0.51	+0.67	+0.58	+0.55	+0.61	+0.69	+0.49	+0.42	+0.34	+0.47	+0.36	+0.37	+0.34	+0.24
Alpha-ReQ	—	—	—	—	-0.00	+0.03	-0.46	-0.40	-0.58	-0.57	-0.35	-0.27	-0.19	-0.26	-0.22	-0.20	-0.28	-0.28
Effective rank	—	—	—	—	+0.02	+0.27	+0.69	+0.63	+0.80	+0.74	+0.57	+0.48	+0.45	+0.50	+0.45	+0.48	+0.48	+0.32
Anisotropy	—	—	—	—	-0.12	+0.17	-0.24	-0.18	-0.43	-0.29	+0.00	-0.01	+0.01	+0.04	+0.08	+0.04	+0.04	-0.20
InfoNCE	—	—	—	—	-0.31	-0.74	-0.77	-0.81	-0.50	-0.76	-0.79	-0.63	-0.66	-0.72	-0.65	-0.72	-0.56	-0.43
LiDAR	—	—	—	—	-0.51	-0.85	-0.22	-0.30	-0.51	-0.31	-0.42	-0.31	-0.34	-0.36	-0.37	-0.35	-0.25	-0.24
Curvature	—	—	—	—	+0.01	+0.11	-0.21	-0.28	+0.05	-0.14	-0.15	-0.06	-0.14	-0.08	-0.12	-0.16	-0.07	+0.02
PTE	—	—	—	—	+0.55	+0.91	+0.35	+0.59	+0.48	+0.52	+0.50	+0.43	+0.46	+0.61	+0.50	+0.56	+0.42	+0.45

Metric	Chords MIREX	Chords root	Chords thirds	Chords triads	Key	Pitch	Beat	Downbeat	Instrument	Emotion	Genre	Structure	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50	MTT	DimSim
ID (GRIDE k=8)	—	—	—	—	-0.57	+0.19	—	—	+0.90	+0.90	+0.87	+0.77	+0.93	+0.87	+0.88	+0.94	+0.85	+0.49
ID (TwoNN)	+0.23	+0.41	+0.37	+0.29	+0.09	+0.19	—	—	+0.90	+0.78	+0.73	+0.66	+0.86	+0.87	+0.85	+0.94	+0.81	+0.72
Alpha-ReQ	-0.31	-0.48	-0.45	-0.36	+0.02	-0.07	—	—	-0.91	-0.87	-0.76	-0.64	-0.91	-0.96	-0.88	-0.97	-0.92	-0.60
Effective rank	+0.29	+0.47	+0.43	+0.35	-0.01	+0.03	—	—	+0.86	+0.76	+0.79	+0.69	+0.84	+0.97	+0.87	+0.96	+0.88	+0.52
Anisotropy	-0.63	-0.90	-0.83	-0.80	-0.04	-0.31	—	—	-0.49	-0.67	-0.50	-0.55	-0.58	-0.81	-0.53	-0.87	-0.49	-0.41
InfoNCE	-0.35	-0.39	-0.41	-0.42	-0.28	-0.37	—	—	-0.86	-0.84	-0.24	-0.58	-0.87	-0.78	-0.81	-0.83	-0.40	-0.80
LiDAR	—	—	—	—	-0.63	+0.03	—	—	+0.93	+0.92	+0.98	+0.91	+0.96	+0.97	+0.99	+0.97	+0.98	+0.21
Curvature	-0.17	-0.12	-0.17	-0.19	-0.25	—	—	—	+0.03	+0.23	+0.09	+0.11	+0.17	—	+0.13	—	+0.47	+0.08
PTE	+0.27	+0.56	+0.60	+0.55	+0.74	+0.83	—	—	+0.28	+0.19	-0.00	+0.20	+0.16	-0.17	+0.23	-0.25	-0.13	+0.59

Metric	Chords MIREX	Chords root	Chords thirds	Chords triads	Key	Pitch	Beat	Downbeat	Instrument	Emotion	Genre	Structure	MTG-Genre	MTG-Instr	MTG-Mood	MTG-Top50	MTT	DimSim
ID (GRIDE k=8)	—	—	—	—	-0.02	—	—	—	—	-0.10	-0.40	—	—	—	—	—	—	+0.69
ID (TwoNN)	—	—	—	—	-0.12	—	+0.49	-0.34	—	-0.25	-0.49	—	-0.57	-0.51	-0.57	-0.57	-0.55	+0.68
Alpha-ReQ	—	—	—	—	+0.46	—	-0.60	-0.06	—	-0.30	-0.20	—	+0.22	+0.17	+0.19	+0.21	+0.23	+0.06
Effective rank	—	—	—	—	-0.33	—	+0.65	+0.08	—	+0.16	+0.10	—	-0.04	-0.02	-0.05	-0.04	-0.09	+0.17
Anisotropy	—	—	—	—	-0.20	—	-0.04	+0.71	—	+0.69	+0.77	—	+0.94	+0.91	+0.92	+0.94	+0.91	-0.86
InfoNCE	—	—	—	—	+0.32	—	-0.78	-0.01	—	+0.02	+0.25	—	+0.23	+0.16	+0.20	+0.23	+0.17	-0.50
LiDAR	—	—	—	—	-0.64	—	+0.65	+0.83	—	+0.78	+0.83	—	+0.69	+0.70	+0.74	+0.70	+0.69	-0.71
Curvature	—	—	—	—	+0.34	—	—	—	—	-0.25	+0.25	—	—	—	—	—	—	-0.40
PTE	—	—	—	—	+0.77	—	-0.86	-0.75	—	-0.50	-0.39	—	-0.40	-0.41	-0.51	-0.43	-0.40	+0.35

Metrics use the paper’s conventions: curvature in the signed convention (monotone to the stored absolute values), and PTE defined as negated phase error with the published rule of taking the strongest of its four probe × error configurations per model and task — so positive always means the metric tracks better layers. Chord recognition is shown per mir_eval metric because the story differs by metric. Correlations here cover all models with data — more than the paper’s figure. Raw profiles: data/.