Melo: Production LLM Music Recommendation Agent¶

Core mechanism evidence. Paper and local evaluation are separate.

Paper information¶

Field	Value
Paper	arXiv 2607.23718
Affiliations	NetEase Cloud Music / Zhejiang University of Technology
First public release	2026-07-26 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`melo`
Local code	`src/auto_research/reproductions/melo/`

Original paper¶

Background and contribution¶

Orchestrate intent, retrieval, recommendations, and explanations across agent nodes. Catalog grounding prevents unsupported entities; failures trigger reflection and retry.

flowchart LR
  Q[Request] --> I[Intent routing] --> C[Grounded catalog retrieval] --> V[Validate and explain]
  V -->|Failure| R[Reflect and retry] --> C

Key figure from the paper¶

Original Figure 2: Melo. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ \hat e=\arg\max_{e\in\mathcal E}s(q,e),\quad \text{retry}=\mathbf1[\hat e\notin\mathcal E\ \lor\ score<\tau]. \]

Paper results¶

System A/B on roughly one million users, 2026-04-02 through 05-10, reports >2 pp playlist retention. Includes Muse Mix UI/product effects.

Local reproduction¶

MovieLens executes grounding, validation, and one retry with repair traces. Against no repair: Hit@10 −20.00%, NDCG@10 −8.16%, fresh Hit@10 +50.00%. Metrics.

auto-research reproduce --paper melo --data-root data --seed 42

Reproduction limits¶

Not a production reproduction. Model, product, and UI effects cannot be isolated and no confidence intervals were reported.