ByteDance's MuChator boosts Douyin Music active days by 46% with conversational AI

ByteDance has deployed MuChator on its Douyin Music platform, a conversational music LLM that transforms passive feed-based discovery into active, intent-driven exploration. Unlike traditional recommendation systems that limit users to algorithm-chosen tracks, MuChator allows users to express situational, colloquial requests like 'play something energetic but not too loud' or 'recommend nostalgic 90s rock'. The system is built on three core innovations: first, a three-stage music knowledge pre-training scheme that increments injects objective music knowledge, subjective knowledge, and personalized preferences into the LLM. Second, context-aware instruction tuning uses an automated pipeline to generate high-quality user-query-music triplets, aligning the model with vague or underspecified intents. Third, hybrid reward modeling (Hybrid RM) jointly optimizes relevance, personalization, and constraints using GRPO-based reinforcement learning.

In extensive evaluations on industrial music recommendation datasets, MuChator outperformed leading proprietary models such as Gemini-3-Pro. More importantly, an online A/B test on millions of Douyin Music users showed a 46.49% improvement in user active days, proving the practical value of conversational music discovery. This approach addresses a key limitation of passive recommendation by empowering users to actively shape their listening experience through natural language, making music discovery more proactive, personal, and satisfying.