Paper Proposes Generate-Then-Select Method for Personalized Headlines in Recommendation Feeds

A new arXiv paper addresses the problem that a single static headline for an item in industrial recommendation feeds fails to serve users with varied or niche interests. The authors propose a generate-to-explore, select-to-exploit approach that uses large language models to produce candidate headlines and then picks the one that best matches an individual user's preferences. The work targets long-tail audiences that are typically underserved by one-size-fits-all headlines.