Improving Training Stability for Multitask Ranking Models in Recommender Systems (Google Research, arXiv 2302.09178)

Overview

This forum entry catalogs a Google Research paper published in February 2023:

• Title: Improving Training Stability for Multitask Ranking Models in Recommender Systems
• Authors: Jiaxi Tang, Yoel Drori, Daryl Chang, Maheswaran Sathiamoorthy, Justin Gilmer, Li Wei, et al. (9 authors total)
• Link: https://arxiv.org/abs/2302.09178
• Section: Ranking for Search

Key points

• Multitask ranking models are a standard architecture in industrial recommender systems, sharing representations across multiple objectives (e.g., different types of user engagement).
• Training such models can be unstable because gradients from different tasks may conflict, degrading convergence and final model quality.
• The paper studies this instability and proposes techniques aimed at mitigating inter-task gradient conflict during training.
• The work targets production-scale recommendation settings, where training stability directly affects serving quality.

Caveat about the source post

The body of the original forum post is largely template boilerplate — generic sections on retrieval-augmented generation, benchmark datasets (MS MARCO, BEIR, Natural Questions), LLM-based reranking, and engineering checklists — rather than a faithful summary of this specific paper. Several sections repeat only the paper title where substantive content should be. Readers interested in the actual methods, experimental setup, and quantitative results should refer directly to the arXiv page and the PDF.

Related entries referenced in the post

• Deep Learning to Rank in Industrial Search Engines
• 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
• Adaptive Neural Ranking Framework: Toward Maximized Business Goal

Practical takeaways

1. Researchers: reproduce the paper's core comparisons and check whether statistical significance and compute cost are reported. 2. Engineers: evaluate how gradient-conflict mitigation techniques integrate with existing multitask training pipelines. 3. Practitioners: weigh offline metric gains against latency, cost, and online satisfaction metrics before adopting training-stability tricks in production.