Technical Summary: FLUID

Problem Statement

Modern recommender systems predominantly rely on ID-based collaborative filtering, where items are represented by unique ID embeddings that accumulate signals from user interactions. However, this paradigm faces a fundamental mismatch in livestreaming recommendation. Unlike long-lived items in e-commerce or short-video domains, a livestream room typically broadcasts for only ~45 minutes before becoming obsolete. Consequently, the item ID embedding remains in a persistent cold-start state, failing to converge within the room's short lifetime.

This creates two specific challenges:

1. Signal Scarcity: The ID embedding table cannot memorize sufficient collaborative signals before the room ends, rendering the ID useless for generalization.
2. ID-Dominance Effect: In existing industrial systems, multimodal features are often added as auxiliary signals alongside the item ID. However, ranking models tend to default to the ID signal (memorization) and underutilize multimodal inputs. In livestreaming, where the ID signal is ephemeral and noisy, this "ID-dominance" wastes the model's capacity on a rapidly decaying signal, preventing the multimodal features from effectively guiding recommendations.

Methodology: The FLUID Framework

FLUID (Framework for Live Universal ID-free Recommendation) is the first framework to fully retire the candidate-side item ID from a production-scale livestreaming ranker, replacing it entirely with multimodal semantic codes. The pipeline consists of four stages:

1. Cross-Domain Multimodal Encoder

To address the sparsity of live data, FLUID employs a single-tower multimodal encoder jointly trained on both short videos and livestreams.

• Architecture: It combines a vision module (SigLIP2-base ViT), a text module (tokenizing titles, OCR, ASR, comments, etc.), and a fusion module based on Qwen3-Embedding-0.6B.
• Training: The encoder uses a query-to-item (Q2I) contrastive objective with InfoNCE loss. It utilizes a two-stage training recipe: (1) Alignment (training only projectors) to preserve pre-trained semantics, followed by (2) Joint fine-tuning.
• Goal: This cross-domain transfer leverages the denser, less noisy supervision from short videos to produce robust semantic representations for the sparse live domain.

2. Discrete Representation (LUCID)

The continuous 128-d embedding ($z$) from the encoder is discretized into LUCID (Live Universal Content IDentifier) codes using Residual Quantization K-Means (RQ-KMeans).

• Structure: LUCID is a hierarchical 4-level codeword tuple ($c_1, c_2, c_3, c_4$).
• Granularity:
  • Slice-level LUCID: Captures 2-minute content dynamics (e.g., a streamer switching from chatting to singing).
  • Room-level LUCID: Obtained via level-wise majority voting over all slices in a session. This provides a stable, persistent identity for the room, decoupling the identifier from transient content changes.

3. Prefix N-gram Embedding

To map the LUCID tuples into the ranker, FLUID uses a prefix n-gram embedding scheme rather than simple level-wise decoding.

• Mechanism: Each level's embedding is conditioned on the full prefix path ($\bar{c}_l = \sum c_k \cdot N^{l-k}$). This ensures that the same codeword at a deeper level represents different semantic regions depending on its parent path, resolving the ambiguity inherent in residual quantization.
• Implementation: Slice-level and room-level LUCIDs use separate embedding tables to prevent the mixing of stable room semantics with transient slice dynamics.

4. Late-Fusion ID-Free Ranker & Staged Warmup

The final architecture replaces the item ID token entirely with LUCID tokens.

• Late Fusion: Slice and Room LUCIDs are injected as independent tokens into the transformer-based ranker backbone. This allows the model to learn interactions between the stable room identity and transient slice dynamics without forcing them into a single representation.
• Staged Warmup: To stabilize the transition from ID-based to ID-free training in an online incremental setting, FLUID employs a three-stage warmup:
  1. Slice Add-on: Add slice LUCID as a new token alongside the converged item ID.
  2. Item ID Phase-out: Progressively zero out the item ID, forcing the model to rely on slice LUCID.
  3. Room Add-on: Introduce room-level LUCID (initialized from the warmed slice tables) to assume the role of persistent identity.

Key Contributions

1. Problem Framing: The paper argues that for ephemeral items, the "ID-dominance" effect is a fundamental bottleneck, not a tolerable nuisance. It posits that the solution is not better fusion tricks, but the full retirement of the candidate-side item ID.
2. Cross-Domain Encoder: Introduction of a multimodal encoder trained jointly on short videos and livestreams to produce hierarchical LUCID codes, serving as the sole candidate-side content identifier.
3. ID-Free Architecture: A late-fusion ranker design where LUCID supplants the item ID, stabilized by a staged warmup procedure that ensures safe transition during online training.

Experimental Results

FLUID was deployed on industrial livestreaming platforms with a combined user base of over 1 billion.

Offline Metrics (CTR AUC):

• Adding slice LUCID to the baseline (with ID) yielded a +0.22% AUC gain.
• Removing the item ID without a substitute caused a significant drop (-0.47%), confirming the model's reliance on ID memorization.
• The final FLUID configuration (No ID + Slice & Room LUCID) achieved a +0.23% gain over the baseline.

Online A/B Test Results:
Compared to the production baseline, FLUID delivered significant gains across multiple metrics:

• Engagement Quality: +0.55% Quality Watch Duration, +0.51% Quality Watch Session.
• Cold-Start & Diversity: +2.05% Cold-Start Room Views, +2.87% Niche Room Views, +1.63% Unique Watched Tags.
• Retention: +0.05% Active Hours (the only configuration to achieve positive retention gains; removing the ID without LUCID substitution resulted in negative retention).

Ablation Insights:

• Fusion: Early fusion (merging ID and LUCID) failed to improve performance, as the model collapsed back to relying on the ID. Late fusion was essential.
• Training: Naive training of LUCID without staged warmup failed. The staged warmup was critical to prevent the model from collapsing onto the ID signal during the transition.
• Embedding: Prefix n-gram embeddings outperformed level-wise decoding, and independent tables for slice/room codes outperformed shared tables.

Significance and Claims

The paper claims that FLUID demonstrates a new principle for short-lived recommendation domains: when items are inherently ephemeral, retiring the item ID is a more principled remedy than attempting to fuse multimodal signals with a decaying ID.

By successfully replacing the ID with multimodal semantic codes (LUCID) and stabilizing the transition via staged warmup, FLUID achieves:

1. Restored Matching Quality: The model can effectively match users to content without relying on insufficient ID history.
2. Genuine Diversity: Improvements in niche and cold-start views are driven by better content understanding, not just the passive redistribution of traffic due to poor matching.
3. User Retention: The framework converts these diversity and engagement gains into actual user retention, proving that content-grounded ranking can outperform ID-based memorization in the livestreaming context.

The work positions itself as a step toward content-grounded ranking in domains where item lifetimes are too short for traditional collaborative filtering to function.