Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning User Interests via Reasoning and Distillation for Cross-Domain News Recommendation

Mengdan Zhu, Yufan Zhao, Tao Di, Yulan Yan, Liang Zhao

TL;DR

The paper tackles cross-domain news recommendation by inferring reusable user interests through a reasoning-driven reinforcement learning framework that trains instruction-tuned LLMs to generate query lists. It introduces GRPO-based multi-reward optimization to align generated queries with retrieval utility, coverage of distinct interests, and query quality, while examining scaling via model capacity and inference-time sampling. To deploy in production, the authors propose on-policy distillation to transfer a compute-intensive teacher policy to a compact student that preserves performance with low latency. Extensive offline experiments and large-scale online A/B tests demonstrate consistent gains in interest modeling quality and downstream retrieval performance, with notable improvements for cold users and practical deployment benefits on real systems.

Abstract

News recommendation plays a critical role in online news platforms by helping users discover relevant content. Cross-domain news recommendation further requires inferring user's underlying information needs from heterogeneous signals that often extend beyond direct news consumption. A key challenge lies in moving beyond surface-level behaviors to capture deeper, reusable user interests while maintaining scalability in large-scale production systems. In this paper, we present a reinforcement learning framework that trains large language models to generate high-quality lists of interest-driven news search queries from cross-domain user signals. We formulate query-list generation as a policy optimization problem and employ GRPO with multiple reward signals. We systematically study two compute dimensions: inference-time sampling and model capacity, and empirically observe consistent improvements with increased compute that exhibit scaling-like behavior. Finally, we perform on-policy distillation to transfer the learned policy from a large, compute-intensive teacher to a compact student model suitable for scalable deployment. Extensive offline experiments, ablation studies and large-scale online A/B tests in a production news recommendation system demonstrate consistent gains in both interest modeling quality and downstream recommendation performance.

Learning User Interests via Reasoning and Distillation for Cross-Domain News Recommendation

TL;DR

The paper tackles cross-domain news recommendation by inferring reusable user interests through a reasoning-driven reinforcement learning framework that trains instruction-tuned LLMs to generate query lists. It introduces GRPO-based multi-reward optimization to align generated queries with retrieval utility, coverage of distinct interests, and query quality, while examining scaling via model capacity and inference-time sampling. To deploy in production, the authors propose on-policy distillation to transfer a compute-intensive teacher policy to a compact student that preserves performance with low latency. Extensive offline experiments and large-scale online A/B tests demonstrate consistent gains in interest modeling quality and downstream retrieval performance, with notable improvements for cold users and practical deployment benefits on real systems.

Abstract

News recommendation plays a critical role in online news platforms by helping users discover relevant content. Cross-domain news recommendation further requires inferring user's underlying information needs from heterogeneous signals that often extend beyond direct news consumption. A key challenge lies in moving beyond surface-level behaviors to capture deeper, reusable user interests while maintaining scalability in large-scale production systems. In this paper, we present a reinforcement learning framework that trains large language models to generate high-quality lists of interest-driven news search queries from cross-domain user signals. We formulate query-list generation as a policy optimization problem and employ GRPO with multiple reward signals. We systematically study two compute dimensions: inference-time sampling and model capacity, and empirically observe consistent improvements with increased compute that exhibit scaling-like behavior. Finally, we perform on-policy distillation to transfer the learned policy from a large, compute-intensive teacher to a compact student model suitable for scalable deployment. Extensive offline experiments, ablation studies and large-scale online A/B tests in a production news recommendation system demonstrate consistent gains in both interest modeling quality and downstream recommendation performance.
Paper Structure (26 sections, 12 equations, 3 figures, 6 tables)

This paper contains 26 sections, 12 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. System Overview
  4. Method
  5. Problem Formulation
  6. Noise Cleaning
  7. Interest Generation with Multi-Reward Reinforcement Learning
  8. Policy Formulation.
  9. Reward Design.
  10. On-Policy Distillation of Interest Models
  11. Experiments
  12. Experimental Setup
  13. Datasets.
  14. Evaluation Metrics.
  15. Baselines.
  16. ...and 11 more sections

Figures (3)

  • Figure 1: Overview of the Reasoning-driven User Interest Generation for News Search Queries (Teacher Model) and On-policy Distillation for Online Serving (Student Model) in the large-scale RecSys.
  • Figure 2: Model capacity scaling for interest generation. All models belong to the Qwen2.5 family and are trained with identical objectives and inference settings. Increasing parameter size from 0.5B to 32B consistently improves Recall@10 and NDCG@10, demonstrating scaling-like behavior in deep user interest modeling.
  • Figure 3: Inference-time sampling scaling with best-of-$N$ selection. Increasing the number of sampled candidate interest lists ($N \in \{1,2,4,8,16,32\}$) consistently improves Recall@10 and NDCG@10, exhibiting diminishing returns at larger $N$.