MuChator: Conversational music discovery¶
Core public-data proxy, not private production dialogue.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2605.27103 |
| Institution | ByteDance. |
| First public release | 2026-05-26 (arXiv v1) |
| Author code | No public author code found as of 2026-09-02. |
| Local adapter or method key | muchator |
| Local reproduction code | src/auto_research/reproductions/muchator/ |
Paper method and evidence¶
Continue pretraining, supervised fine-tuning and reinforcement learning in music so the conversational model can infer current intent and actively discover options instead of merely repeating past preferences.
flowchart LR
U[Current user expression] --> I[Infer music intent]
I --> D[Propose discovery options]
D --> R[Intent, discovery and dialogue reward]Original paper figure¶
The original figure shows the discovery and training path. From the original paper; copyright remains with the authors.
Core equation¶
\[
\pi^*=\arg\max_\pi\\;\mathbb E[r_{intent}+r_{discovery}+r_{dialogue}]-\beta D_{KL}(\pi\Vert\pi_{ref}).
\]
Reported results¶
About one month of Douyin Music A/B reports active days +46.49%, usage time +77.36%, and CTR +11.26%. These are paper online metrics only.
Local reproduction¶
Same public MovieLens-1M full-catalog protocol: NDCG@10 0.05401→0.04477 (-17.09%). Three-seed results, mean, SD, 95% CI. Proxy objectives do not reproduce private logs, bucketing, model scale or serving.