📝 Paper Summary
News Recommendation
Diversity Nudges
LLM-based Personalization
The paper introduces a news recommendation system that combines algorithmic dual-calibration (balancing topic and locality) with LLM-rewritten headlines to nudge users toward consuming diverse domestic and global news.
Core Problem
Recommendation systems often optimize for short-term engagement, creating filter bubbles, while users frequently ignore diverse content even when it is exposed to them (the exposure-consumption gap).
Why it matters:
- Reinforcing narrow preferences prevents the long-term societal goal of a well-informed public
- Existing research focuses on exposure diversity (showing diverse items) but fails to convert this into actual consumption (users clicking diverse items)
- Users find diverse content cognitively demanding or irrelevant if not framed correctly
Concrete Example:
A user who primarily reads U.S. sports news might be algorithmically pigeonholed into a 'Domestic Sports' bubble. Even if a 'World Politics' article is recommended for diversity, the user ignores it because the headline seems irrelevant. The proposed system would rewrite the World news headline to highlight a connection to a domestic event the user previously read.
Key Novelty
Topic-Locality Dual Calibration + LLM Relevance Nudges
- Extends calibration beyond just topics (e.g., Sports vs. Politics) to include locality (Domestic vs. World), ensuring geographic balance within topics
- Uses Large Language Models (LLMs) to rewrite news previews (headlines/subheads) for diverse articles, explicitly explaining their relevance to the user's reading history to reduce cognitive friction
Architecture
The pipeline for generating personalized news previews for diverse articles
Breakthrough Assessment
7/10
Novel combination of calibration objectives and generative UI nudges tested in a real-world longitudinal study (POPROX), addressing the critical gap between exposure and consumption diversity.