Reproduce “MuChator: Enabling Active Music Discovery via Conversational Music LLMs in Douyin Music” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiahao Liang, Linzhi Huang, Xuannan Liu, Xukai Wang, Xuanpu Luo, Yongchun Zhu, Jingwu Chen, Feng Zhang and 1 more
MuChator, an LLM-based framework from ByteDance, enables active music discovery through conversational interaction on Douyin Music. It systematically integrates music-domain knowledge, collaborative reasoning, and personalized preferences, leading to significant online improvements like a 46.49% increase in User Active Day and a 77.36% increase in playback Duration.
Key insights
- A multi-stage curriculum pre-training strategy effectively bridges the distribution gap between general LLM corpora and specific music recommendation data, progressively building domain knowledge and collaborative reasoning.
- Interleaving item and feedback prediction within a unified autoregressive sequence for 'next-behavior prediction' during pre-training provides denser supervision and enhances the modeling of personalized user preferences.
- A hybrid reward function, integrating model-based relevance and personalization signals with rule-based constraints, is critical for fine-grained alignment of LLMs with complex user intents, personalized tastes, and practical recommendation requirements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.