Researchers have developed NEXT, an 8-billion parameter vision-language model designed for video recommendation. NEXT analyzes a user's recently watched video to infer their next interest and then retrieves relevant follow-up content. The model was trained using a three-stage process including reinforcement learning and supervised fine-tuning, achieving strong performance on visual question answering tasks and outperforming larger models. When deployed in a large-scale social media system, NEXT demonstrated significant improvements in watch time and distinct video exposure. AI
IMPACT This model could improve user engagement and content discovery in video platforms by providing more relevant recommendations.
RANK_REASON Publication of a research paper detailing a new vision-language model and its application. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
