- The paper introduces FLUID, the first ranking system that fully retires candidate-side item IDs for livestream recommendation.
- It employs a cross-domain multimodal encoder and RQ-KMeans to generate hierarchical semantic codes (LUCID) from transient livestream content.
- A staged warmup protocol transitions from ID reliance to content-grounded ranking, yielding significant gains in engagement, cold-start views, and retention.
FLUID: Ephemeral IDs to Multimodal Semantic Codes in Livestream Recommendation
Introduction
ID-based collaborative filtering underpins modern industrial recommendation systems, leveraging long-lived item IDs to accumulate implicit and explicit user feedback through persistent dense embedding tables. Livestreaming significantly departs from this paradigm. The intrinsic short lifetime of live rooms—averaging approximately 45 minutes—prevents the formation of meaningful item representation within the model's embedding space. As a result, most live room IDs spend their entire existence in cold-start, causing large recommendation models (LRMs) to collapse into memorizing transient, non-generalizable signals. The ID-dominance effect, tolerable in video or e-commerce domains, becomes a fundamental bottleneck in livestreaming, both practically and theoretically.
This work introduces FLUID (Framework for Live Universal ID-free Recommendation), which is the first ranking architecture in industrial deployment to fully retire the candidate-side item ID. Rather than continuing to bolster ephemeral IDs with additional features, FLUID constructs a fully content-grounded representation using a cross-domain multimodal encoder, discretized into hierarchical semantic codes (LUCID), and integrates these codes into the production ranker via late fusion, stabilized by a dedicated staged warmup protocol. The paper reports statistically significant online gains not only in engagement and cold-start recommendation, but also in retention and content diversity, across over a billion active users.
Problem Analysis
The mismatch between short-lived livestreaming items and ID-centric architectures leads to poorly-trained embeddings and persistent cold-start. Conventional industrial approaches augment the item ID with content-derived features (multimodal, textual, or metadata), either as auxiliary dense inputs or via explicit alignment/discretization into semantic codes. However, in these architectures, multimodal features serve as secondary information, underexploited due to strong model reliance on the simultaneously-present ID. Several prior works (e.g., [rajput2023recommender], [liu2024alignrec], [qarm2024]) demonstrate this "ID-dominance" empirically, with the ranking model failing to utilize content-based signals if an item ID is present—even when the ID itself is poorly generalizable.
For livestreaming, these limitations are exacerbated by the multi-modal, highly dynamic content (e.g., quick transitions from chatting to singing, rapidly evolving audience interaction, and more), challenging the computation of robust, generalizable semantic representations. Single-domain encoders, trained solely on live data, cannot fully leverage collaborative patterns due to label sparsity and noise.
FLUID Methodology
Cross-Domain Multimodal Encoder
FLUID's encoder is a single-tower transformer jointly trained across both short video and livestream content domains. Inputs include visual representations (using SigLIP2-based ViT for image/video data), and rich, multi-source text (titles, OCR, ASR, bios, comments, tags), fused via Qwen3-Embedding. Training employs a contrastive InfoNCE objective on query-to-item pairs, where queries are user search terms or MLLM-synthesized keywords, and positives are user interactions (likes, watches, shares). Two-stage training—aligning projector/projection first, then full model fine-tuning—preserves pre-trained LLM capabilities and facilitates cross-domain semantic transfer.
Discrete Code Generation with RQ-KMeans
The encoder yields a 128-d embedding per 2-minute session slice, which is then discretized through multi-level residual quantization (RQ-KMeans, L=4, N=64), producing a hierarchical LUCID code per slice. A room-level LUCID is computed by per-level majority-vote over all a room's slices, yielding a stable identifier suitable for aggregation of longer-range signals.
Prefix N-Gram Embedding
To embed LUCID codes, FLUID adopts a prefix n-gram strategy, resolving the semantic ambiguity inherent in residual quantization. Each possible codeword prefix path is mapped to a unique slot (akin to tree traversal), and embedding lookup is performed on the composite path. Separate tables for slice-level and room-level LUCIDs prevent parameter interference between transient and persistent representations.
Late Fusion and Staged Warmup
FLUID eliminates the candidate-side item ID, replacing it with room-level and slice-level LUCID tokens injected independently into the ranker backbone. The ranking network is a transformer-based architecture with token-level feature interaction.
Transitioning away from the ID-based system is managed via a three-stage curriculum:
- Slice Add-on: Introduce slice LUCID as a new candidate token, keeping item ID.
- ID Phase-out: Gradually ablate the candidate-side ID, forcing the model to rely solely on LUCID.
- Room Add-on: Instantiate separate room-level LUCID embeddings, transferring weights from pre-trained slice tables, and train end-to-end.
This protocol prevents model collapse and ensures stable optimization, overcoming the intrinsic optimization asymmetry between memorization (ID) and generalization (content).
Empirical Results
FLUID is deployed in the production pipeline of global-scale livestreaming platforms (over 1 billion users). The final deployed architecture, with both slice and room LUCID and no candidate ID, achieves the following absolute improvements in longitudinal online A/B tests:
- +0.55% Quality Watch Duration
- +2.05% Cold-Start Room Views
- +2.87% Niche Room Views
- +1.63% Unique Watched Tags
- +0.05% Active Hours
Ablative benchmarking demonstrates that supplementing item ID with LUCID provides only marginal benefit, as the model continues to ignore content features in favor of memorization. Crucially, removing the candidate-side item ID without any content substitute is strongly detrimental to short-term engagement and retention, and the apparent diversity improvement reflects traffic randomization rather than effective broadening. By contrast, the full FLUID configuration restores or surpasses pre-removal engagement while increasing exposure diversity and user retention.
On both CTR AUC and semantic retrieval/classification metrics, the single-tower multimodal architecture, cross-domain pretraining, prefix n-gram embeddings, and separate slice/room tables each produce positive additive effects.
Theoretical and Practical Implications
The results establish a rigorous principle for ephemeral-content recommendation: when candidate item IDs are inherently non-persistent, retiring candidate-side ID embeddings is superior to further content-ID fusion or hybridization. FLUID recasts the core matching paradigm: persistent IDs are replaced with semantically stable, content-derived codes at multiple granularity levels, learned via multimodal transfer and discretized for compatibility with large-scale recommenders.
This approach generalizes beyond livestreaming to any recommendation setting with ephemeral inventory—e.g., breaking news, social audio/video rooms, and certain ad channels. Architecturally, FLUID bridges the gap between LLM-driven multimodal representation and fully content-grounded industrial ranking. Its staged optimization protocol may have broader applicability to other transitions in large-scale models where memorization and generalization signals are in tension.
Future Developments
Potential extensions include the expansion of LUCID's dynamic range (more quantization levels or residual depths), integration with generative multimodal retrieval, and the application of similar ID-free modeling in domains with rapidly shifting catalogs or user-item graphs. Advances in universal multimodal representation learning and efficient online incremental training can further improve FLUID's generalization and latency. Cross-domain shared vocabularies for content codes may facilitate unifying recommendation across multiple modalities and verticals.
Conclusion
FLUID demonstrates that in environments with fundamentally short-lived content inventory, the canonical ID-dominant recommendation architecture is suboptimal. By fully retiring candidate-side item IDs and grounding recommendation in hierarchical, cross-modal, content-derived semantics, FLUID advances both the theory and practice of industrial-scale recommendation under highly dynamic conditions. The proposed methodology is substantiated by industrial-scale gains in engagement, coverage, and retention, and provides a robust framework for content-grounded, multimodal recommendation in ephemeral item scenarios (2605.21832).
