The Bregman Cue-Sensitivity Map: Auditory-Scene-Analysis Cues Across Music Foundation Models

We ask which of four classical auditory-scene-analysis (ASA) cues — frequency proximity, temporal proximity, harmonicity, onset synchrony — the internal representations of released music foundation models are sensitive to, and how that depends on paradigm and scale. We define the Bregman Cue-Sensitivity Score (BCS), a rank correlation within pre-specified anchor clusters, averaged over a common relative-depth grid, between a model's representational distances and an ASA-parameterised stimulus axis, and compute it for twelve released models — self-supervised encoders, autoregressive (AR) and masked non-autoregressive (NAR) decoders, diffusion transformers (DiT) — over 8,530 rendered stimuli. Across the sampled checkpoints the result is a family-structured map rather than a simple scale ladder (S = 0.782; lineage-blocked permutation p = 0.019): encoders track frequency and onset almost fully (ρ ≥ 0.92) but stay modest on temporal proximity, whereas the larger AR decoders invert this, temporal-proximity BCS rising from 0.055 to 0.812 across the MusicGen 300M-to-3.3B ladder (Δρ = +0.757, 95% CI [0.69, 0.82], Holm p = 0.001). Two layers of front-end control then relocate the informative part of the map: frequency, harmonicity and onset are reproduced without learning, by fixed log-mel, constant-Q and gammatone transforms and by the six learned codec/VAE front-ends the models actually read, whereas temporal proximity is the one cue none of the nine reproduces (largest |ρ| of 0.136) and the one on which 7 of 9 transformers resolve above their own front-end; matched random initialisation over three seeds accounts for about a third of its rise with scale. All statements are observational and sample-scoped.