TubiFM: Unified Item, Carousel, and Search Ranking for Streaming Discovery

Abstract: Personalized discovery systems often train separate models for item ranking, carousel ranking, and search, even though these tasks expose complementary signals from the same viewer journey: watches shape carousel and item ranking, search queries reveal intent even when they do not lead to a catalog match, and watch history helps interpret search as rewatching, continuation, or new discovery. We introduce the user story, a serialized representation that turns a user's cross-surface history - attributes, sessions, watch events with surface and carousel context, and search events - into a single token sequence. By interleaving pretrained language tokens with domain-specific event tokens, user stories let heterogeneous recommendation and search tasks be expressed as prompted next-token prediction over a shared grammar. TubiFM is one instantiation of this approach: a Llama 3.2 1B-based model trained on user stories and prompted to rank items, carousels, or search results without task-specific architectures. In offline evaluation, this single model outperforms specialist baselines across item, carousel, and search ranking. In online A/B tests, TubiFM significantly improves search total viewing time (TVT) by $+3.9\%$ and carousel TVT by $+0.30\%$. Item ranking is statistically neutral on TVT ($+0.14\%$), but matches a mature production stack; across all three tasks, TubiFM serves on L40S GPUs and reduces p99 ranking latency from 500ms to 200ms. These results show that shared user stories can improve discovery while simplifying ranking systems.

• The paper introduces a unified ranking model using serialized user stories that consolidates item, carousel, and search ranking with a finetuned Llama 3.2 model.
• It demonstrates significant improvements in HR@8 and NDCG@8 over task-specific baselines through effective cross-task signal synergy.
• Practical evaluations reveal reduced ranking latency and simplified serving architecture, highlighting the model's industrial applicability.

Unified Cross-Surface Ranking via Serialized User Stories: An Analysis of "TubiFM: Unified Item, Carousel, and Search Ranking for Streaming Discovery" (2605.23702)

Introduction

The paper presents TubiFM, a unified foundation-model-style ranking system for streaming discovery, capable of item, carousel, and search ranking within a single architecture. The approach centers on the concept of the "user story"—a serialized, hierarchical token sequence capturing the user’s cross-surface journey through watches, carousels, searches, along with contextual and temporal information. TubiFM instantiates this paradigm using a finetuned Llama 3.2 1B model that leverages prompted inference over the shared grammar of user stories, eliminating the need for task-specific models or feature engineering traditionally required in production recommendation systems.

User Story Representation

The user story is devised as a granular, typed sequence capturing user attributes, session demarcations, and ordered events. This includes:

• User attributes: Coarse features (e.g., device, country), serialized as ordinary text, enabling extensibility.
• Sessions: Explicit session delimiting tokens encoding inactivity and recency, plus associated temporal features (elapsed-time, day of week).
• Events:
  • Watch events: Item ID, duration, surface, carousel, timestamp.
  • Search events: Query tokens and timestamp, including intermediate states and outcomes.

The serialized user story format enables direct autoregressive sequence modeling and simplifies adding new features by token insertion, rather than explicit architectural modifications.

TubiFM Model Architecture and Training

TubiFM is instantiated as a finetuned Llama 3.2 1B model with an extended tokenizer incorporating domain-specific event and item tokens. Training is conducted on approximately 20M user stories, each capped at 1024 tokens. The auxiliary catalog corpus bridges domain tokens to natural language, enhancing semantic alignment especially for search-related tasks. The training objective is next-token prediction across mixed domains; stochastic masking during training enables container-independent ranking and robust catalog refresh handling.

All ranking tasks—item, carousel, and search—are postulated as next-token prediction under variable prompt heads. Task type is controlled via prompt suffixes, allowing the model to operate in unified or task-adaptive regimes.

Offline and Online Evaluation

Offline Benchmarks

Using Hit Rate (HR@K) and NDCG@K at cutoffs $K \in \{8, 50, 100\}$, the unified TubiFM model demonstrates strong numerical improvements over task-specific baselines:

• Item Ranking:
  • HR@8: 0.5817 vs. 0.4121 (HSTU) — +41.2% relative improvement.
  • NDCG@8: 0.4599 vs. 0.3105 (HSTU) — +48.1% relative improvement.
• Carousel Ranking:
  • HR@8: 0.8343 vs. 0.7724 (HSTU).
  • NDCG@8: 0.6366 vs. 0.5362 (HSTU).
• Search Ranking:
  • HR@8: 0.5673 vs. 0.4637 (BM25).
  • NDCG@8: 0.4480 vs. 0.3732 (BM25).

In all tasks, the unified model surpasses both classical and task-specific TubiFM variants. Notably, unified training boosts HR@8 by 12.1–16.9% over TubiFM finetunes for item and search ranking, affirming the cross-task signal synergy afforded by the serialized stories.

Ablations

Key findings from ablations:

• LLM Initialization: Pretraining via Llama 3.2 checkpoint delivers significant metric improvements across all tasks; the language modeling capacity is directly beneficial despite domain tokens dominating outputs.
• Session Segmentation: Removing session boundaries reduces ranking quality, particularly for item and search tasks; temporal structure is critical.
• User Attributes: Removing attribute tokens marginally degrades metrics, indicating these are weak priors rather than dominant drivers.
• Catalog Masking: Masking strategies enable robust handling of catalog churn and support inference over unseen items without architectural retraining.

Online A/B Tests

Against production systems, TubiFM yields:

• Search: Statistically significant +3.9% TVT lift, with +20% uplift for tail/long queries.
• Carousel: Statistically significant +0.30% TVT lift, especially in top-carousel positions.
• Item: TVT is neutral (+0.14%), demonstrating parity with mature stacks while simplifying serving path.

TubiFM reduces p99 ranking latency from ~500ms to ~200ms across all surfaces, supporting practical deployment under rigorous production constraints.

Implications and Future Directions

The unified user-story modeling approach enables a foundation model to subsume both recommendation and search behaviors, leveraging cross-surface user intent as a composite input. Empirically, the model sets a new standard for ranking quality in streaming discovery, outperforming specialist baselines and commercial stacks. The practical benefits—simplified serving architecture, faster inference, and catalog robustness—are compelling for industrial environments.

Theoretically, serialized user stories offer a reusable schema for diverse recommendation domains with hierarchical, sequential behaviors (e.g., commerce, news, music). The prompt-driven paradigm may generalize to broad multitask personalization, facilitating unified modeling and inference across modalities and surfaces. There is potential for expanding this approach with larger backbone models, cross-domain transfer, and structured search-query intent interpretation, especially as sequential event modeling and generative retrieval architectures mature.

Further research should extend evaluation to domains beyond streaming, investigate label quality (especially negative sampling in search tasks), and explore continual learning regimes for dynamic catalog and behavioral distributions.

Conclusion

TubiFM demonstrates that unified modeling via serialized user stories yields both superior ranking metrics and operational simplicity in streaming discovery. The model’s empirical strength across item, carousel, and search tasks, combined with practical gains in inference latency and serving architecture, positions user-story-based foundation models as an effective strategy for cross-task personalization. The abstraction suggests broader applicability, supporting future work toward unified, prompt-driven recommender systems in heterogeneous sequential domains.