Does model documentation improve the analysis of musical responses?
A planned benchmark of language-model-assisted analysis
Abstract
Can documentation help a language-model analyst draw better conclusions from a computational representation of music? We propose a benchmark using Musical Intelligence (MI), which describes recordings through acoustic and listening-related computations. The main comparison holds numerical arrays and the analyst fixed while adding definitions, temporal support and the record of how quantities were computed. Tasks cover recovery of known relationships, detection of confounds, replication and prediction on held-out data. A successful analysis must produce a supported conclusion from permitted information and numerical results that can be rerun. Matched resources, alternative labels and an expert-written analysis will test competing explanations for any benefit. Musical pleasure is the initial application. This paper describes the framework and planned benchmark; it reports no participant, neural or benchmark results.
Independent future benchmark. Its human-study methods are integrated into the 12 October protocol and supplement. The linked 11 October framework remains available as the earlier proposal. No human results are reported. The current PDF is the version.
Supporting materials
Editable manuscripts and editorial checksZIP ↗Integrated human-study protocolPDF ↗Supplementary Methods · future-study specification in S11PDF ↗Version history
The record stays at this address. Each PDF preserves the version cited.
- Draft · CurrentCite this version ↓
Revised prose and headings. Study specifications, equations, references and figure artwork retained. References follow the text without a forced page break. The cross-reference to the companion study identifies its 12 October amendment.
- DraftCite this version ↓
Inspectable representations and auditable inference in computational music neuroscience
Editorial revision of the abstract, narrative, headings and captions. Figure artwork, equations, references and study specifications are retained. 7 pages; 5 figures. No empirical results are added.
- DraftCite this version ↓
Inspectable representations and auditable inference in computational music neuroscience
Compact edition: 6 main pages plus 1 reference page; 5 combined figures. Methods, numerical specifications, figure content and references retained. No participant results are reported.
- DraftCite this version ↓
Inspectable representations and auditable inference in computational music neuroscience
White-background journal draft with revised typesetting. No participant results. Original PDF, unchanged.