Reproduce “Learning User Interests via Reasoning and Distillation for Cross-Domain News Recommendation” on OpenResearch

About this paper

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

Researchers from Microsoft and Emory University developed a framework for cross-domain news recommendation that employs a reasoning-driven large language model to infer reusable user interests as search queries. Optimized with multi-reward reinforcement learning and distilled into a smaller model for efficient production, this approach improved daily active users by 0.95% and click-through rates by 0.22% in an online A/B test.

Key insights

• Reasoning-driven LLMs, when optimized with multi-reward reinforcement learning, can effectively abstract deeper, reusable user interests from diverse and noisy cross-domain behavioral signals.
• On-policy distillation provides an effective mechanism to transfer complex LLM reasoning capabilities to significantly smaller models, overcoming computational barriers for real-world, large-scale deployment.
• A well-designed composite reward function, incorporating metrics like retrieval alignment and interest coverage, is crucial for guiding LLMs to generate high-quality outputs that simultaneously optimize multiple desirable properties for recommendation systems.

Summary by alphaXiv.