360Brew: A Decoder-Only Foundation Model for Personalized Ranking and Recommendation
Paper: arXiv:2501.16450 — January 2025 Authors: Hamed Firooz, Maziar Sanjabi, Adrian Englhardt, Aman Gupta, Ben Levine, Dre Olgiati, et al. (23 authors) Category: Recommender Engines
Key points
- 360Brew is a decoder-only foundation model designed for personalized ranking and recommendation, replacing multi-stage, task-specific pipeline components with a single unified model.
- User activity history, item metadata, and query context are serialized as text; the model generates rankings and responses directly without dedicated embedding layers or task-specific heads.
- The paper emphasizes end-to-end training on large-scale platform engagement data, enabling the same model to serve multiple surfaces (search ranking, recommendations) rather than maintaining separate models per task.
- Reported results show the approach is competitive with or superior to production baselines on internal ranking and recommendation tasks, while reducing system complexity.
- The work discusses practical deployment considerations: latency and cost budgets, hallucination and safety risks in generative outputs, and the gap between offline metrics and online user satisfaction.
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers (arXiv:2402.17152)
- Augmenting Netflix Search with In-Session Adapted Recommendations (ACM)
- Bridging Language and Items for Retrieval and Recommendation (arXiv:2403.03952)
- Data-efficient Fine-tuning for LLM-based Recommendation (SIGIR 2024)
- DiffKG: Knowledge Graph Diffusion Model for Recommendation (WSDM 2024)
- EAGER-LLM: Enhancing LLMs as Recommenders (arXiv:2502.14735)
- Data: training on behavioral logs requires PII handling, versioned indexes/embeddings, and care with distribution shift between logged and live traffic.
- Latency: single-model ranking must respect p99 budgets; consider cascades, early stopping, and caching for hot queries.
- Evaluation: pair offline ranking metrics (nDCG, Recall@k) with interleaving/online experiments and human audits of generated outputs.
- Safety: generative ranking outputs need filtering, source validation, and bias checks before production rollout.
Context
Traditional recommendation stacks split retrieval, ranking, and generation across specialized components (candidate generators, gradient-boosted or deep CTR rankers, re-rankers). LLM-era systems revisit this division of labor: ranking itself can be framed as a sequence-to-sequence task over serialized behavioral context. 360Brew is an industrial-scale instance of this direction, alongside related work such as action-sequence transducers for sequential recommendation and LLM-based retrieval and recommendation (e.g., Bridging Language and Items for Retrieval and Recommendation, EAGER-LLM).