A Generative Re-ranking Model for List-level Multi-objective Optimization at Taobao (arXiv 2505.07197)

This forum entry indexes a May 2025 arXiv paper from Taobao's recommendation team.

Paper metadata

| Field | Value | |-------|-------| | Title | A Generative Re-ranking Model for List-level Multi-objective Optimization at Taobao | | Authors | Yue Meng, Cheng Guo, Yi Cao, Tong Liu, Bo Zheng | | Published | May 2025 (arXiv) | | Link | https://arxiv.org/abs/2505.07197v1 | | Category | Ranking for Search |

Context

In large-scale search and recommendation systems, re-ranking has traditionally optimized each item independently or relied on cascaded pipelines that treat retrieval, ranking, and presentation as separate stages. This makes it difficult to jointly optimize multiple business objectives (e.g., CTR, conversion, user satisfaction) at the level of the final displayed list.

The paper proposes a generative re-ranking model that operates on the item list as a whole, enabling list-level multi-objective optimization in Taobao's production recommendation stack.

Key points

• List-level optimization: instead of scoring items independently, the model reasons about the entire candidate list, capturing intra-list interactions such as diversity and complementarity.
• Generative formulation: the re-ranking task is framed in a generative paradigm, consistent with the broader Gen-Rec / generative retrieval trend.
• Multi-objective trade-offs: the approach explicitly targets simultaneous optimization of multiple business metrics rather than a single proxy objective.
• Industrial deployment: the work is grounded in Taobao's production environment, where latency, cost, and safety are hard constraints alongside offline metrics.

Positioning within the field

The forum analysis situates the paper in the evolution of learning-to-rank: BM25 → BERT cross-encoders → dense two-tower retrieval → late interaction → generative retrieval and LLM-based agentic search. In recommendation, the trajectory runs from matrix factorization and deep CTR models through sequential Transformers to LLM-driven and generative recommendation (Gen-Rec). Core tensions remain the same: sparse user behavior, huge item catalogs, multi-objective business trade-offs, and the online inference cost of large models.

Related entries

• Deep Learning to Rank in Industrial Search Engines and Recommender Systems
• Multi-Objective Recommendation in the Era of Generative AI: A Survey (arXiv 2506.16893)
• A Thorough Comparison of Cross-Encoders and LLMs for Reranking SPLADE (arXiv 2403.10407)
• Accelerating Listwise Reranking: Reproducing and Enhancing FIRST (SIGIR)
• Adaptive Neural Ranking Framework: Toward Maximized Business Goal

Takeaways

1. Architecture: cascade retrieval + ranking + generation remains mainstream, but generative re-rankers are increasingly treating list composition itself as a learnable decision. 2. Data: high-quality behavioral logs and instruction data remain critical; synthetic data needs safeguards against leakage and distribution shift. 3. Evaluation: the gap between offline metrics and online satisfaction is widening; LLM-as-judge should be cross-validated with human assessment. 4. Deployment: latency, cost, interpretability, and safety are non-negotiable constraints in industrial systems — do not optimize academic benchmarks alone.

> Note: Quantitative results should be verified against the original PDF: https://arxiv.org/abs/2505.07197v1