ELISE: multilingual semantic retrieval for Apple Music¶
Fidelity: core mechanism. Runs the defining mechanism verifiable on public data. Private data, production models, and serving are outside scope.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2607.10239 |
| Organization | Apple |
| First public release | 2026-07-11 (arXiv v1) |
| Authors' code | The paper did not provide official or author code when checked on 2026-09-02. |
| Adapter / method | elise |
| Local reproduction code | src/auto_research/reproductions/elise/ |
Original paper¶
Motivation and method¶
A Siamese bi-encoder complements a lexical inverted index. Quantile distribution matching calibrates dense and token scores to one scale and limits semantic results.
flowchart LR
A["Public behavior and candidates"] --> B["ELISE core mechanism"]
B --> C["Matched-budget method output"]
A --> D["Paper-defined direct baseline"]
C --> E["Shared metrics and structural diagnostics"]
D --> EKey figure from the paper¶
Figure 1: Core architecture, modules, and their connections. From the original paper; copyright remains with the authors.
Core equation¶
Paper results¶
In a global 18-day Apple Music A/B test, conversion improved +2.28%, no-result rate fell -86.0%, and tail-query conversion improved +7.93%. These are paper-reported online measurements only, not local public-data results.
Local reproduction¶
Local comparison: Under the same full-catalog protocol, NDCG@10 is 0.05401 for the baseline and 0.04739 for ELISE, -12.24% relative. This is a negative result on a public proxy task and does not extrapolate to production lift.
See the three-seed artifact for complete results, means, standard deviations, and 95% confidence intervals.
Reproduction boundaries¶
The local run uses a public MovieLens-1M subset and auditable proxy objectives. It does not reproduce private company logs, online traffic assignments, production model scale, or serving. This is core-mechanism validation, not reproduction of the paper's absolute results or online gain.