FLUID: From Ephemeral IDs to Multimodal Semantic Codes for Industrial-Scale Livestreaming Recommendation

Researchers introduced FLUID, a production-scale recommendation system that eliminates reliance on item IDs for livestreaming platforms by using multimodal semantic codes instead. Deployed across platforms with over one billion users, the system achieves significant performance gains including 2.05% improvement in cold-start room views, addressing a fundamental challenge in recommending short-lived broadcast content.

FLUID represents a meaningful advancement in recommendation systems architecture, specifically targeting the cold-start problem that plagues livestreaming platforms where content has extremely short lifespans. Traditional ID-based collaborative filtering accumulates user interaction signals over time, but livestreams typically broadcast for only tens of minutes, leaving their embeddings perpetually undertrained. The FLUID framework solves this by shifting from ID-centric to content-centric characterization using discrete hierarchical semantic codes called LUCID, which are jointly trained across short-video and livestream domains.

The innovation emerged from practical constraints in production recommender systems at massive scale. Rather than waiting for ID embeddings to mature, FLUID uses a staged warmup approach: first introducing cold, slice-level LUCID codes as independent tokens alongside ID embeddings, then progressively replacing ID embeddings with room-level LUCID before online training. This methodology reflects a broader industry trend toward multimodal learning and content-understanding approaches that can generalize across cold-start scenarios.

The deployment metrics demonstrate tangible business impact. A 2.05% improvement in cold-start room views directly addresses a critical monetization challenge for livestreaming platforms, where viewer acquisition for new broadcasts drives platform growth. The 0.55% quality watch duration gain and 0.05% active hours increase indicate the system maintains engagement quality while expanding reach. For platform operators and ML engineers, FLUID offers a reproducible blueprint for handling ephemeral content at industrial scale.

Looking forward, similar semantic-code approaches may migrate into other recommendation domains facing cold-start challenges, including emerging creator content and niche categories. The cross-domain training methodology also suggests potential applications in personalized content discovery systems beyond livestreaming.

• →FLUID eliminates item ID dependency in livestreaming recommendation by using multimodal semantic codes (LUCID) trained jointly on short videos and livestreams.
• →The system achieves 2.05% improvement in cold-start room views through a staged warmup approach that progressively replaces ID embeddings.
• →Deployed across billion-user platforms, FLUID delivers measurable gains: +0.55% watch quality duration and +0.05% active hours improvements.
• →Content-centric recommendation systems show superiority over ID-based approaches for ephemeral content with short broadcast windows.
• →Cross-domain multimodal training enables better generalization for cold-start scenarios in recommendation systems.