RankTower: A Synergistic Framework for Enhancing Two-Tower Pre-Ranking Models
Overview
This post introduces RankTower, an academic paper published on arXiv in July 2024 (arXiv:2407.12385) by YaChen Yan and Liubo Li.
| Field | Content | |------|------| | Title | RankTower: A Synergistic Framework for Enhancing Two-Tower Pre-Ranking Model | | Authors / Affiliation | YaChen Yan, Liubo Li | | Published | July 2024 (arXiv) | | Source | https://arxiv.org/abs/2407.12385 | | Category | Ranking for Search |
Background and Motivation
In large-scale search, recommendation, and personalization systems, ranking has long faced challenges around efficiency, scalability, and user intent understanding. Traditional pipeline-based approaches often treat retrieval, ranking, and generation as disconnected stages. Pre-ranking — the stage between retrieval and fine ranking — typically relies on two-tower architectures to achieve low-latency scoring at very high throughput, trading off some model expressiveness for speed.
RankTower addresses this setting by proposing a synergistic framework for enhancing the two-tower pre-ranking paradigm, aiming to improve ranking quality without sacrificing the latency and scalability guarantees required by industrial systems.
Context Within the Ranking Literature
Neural ranking has evolved from BM25 to BERT-style cross-encoders, two-tower dense retrieval, late-interaction models, and generative retrieval. Each generation balances the triangle of efficiency, effectiveness, and maintainability:
- Two-tower dense models enable millisecond-level recall via approximate nearest-neighbor search, but can be sensitive to domain shift and long-tail queries.
- Cross-encoders achieve higher accuracy but cannot pre-compute document representations.
- Pre-ranking models sit in the middle of the cascade, filtering candidates before fine ranking, so even small quality gains here propagate through the entire system.
- Deep Learning to Rank in Industrial Search Engines, Recommender Systems (survey)
- Multi-Objective Recommendation in the Era of Generative AI: A Survey
- A Generative Re-ranking Model for List-level Multi-objective Optimization
- A Thorough Comparison of Cross-Encoders and LLMs for Reranking SPLADE
- Accelerating Listwise Reranking: Reproducing and Enhancing FIRST (SIGIR)
- Adaptive Neural Ranking Framework: Toward Maximized Business Goal
- Original paper: RankTower: A Synergistic Framework for Enhancing Two-Tower Pre-Ranking Model. arXiv, July 2024. https://arxiv.org/abs/2407.12385
Related Entries
Takeaways for Practitioners
1. Architecture: The cascade of retrieval + pre-ranking + fine ranking + re-ranking remains the industrial mainstream; pre-ranking quality directly bounds downstream performance. 2. Latency budgets: p99 constraints, caching strategies, and early-stopping remain hard engineering constraints that paper metrics alone do not capture. 3. Evaluation: Offline ranking metrics (nDCG, Recall@k) should be validated against online A/B results such as CTR and user satisfaction. 4. Reproducibility: Quantitative results in this post are not reproduced here — please verify specific numbers against the original arXiv PDF.