Netflix Artwork Personalization via LLM Post-training скачать в хорошем качестве

Netflix Artwork Personalization via LLM Post-training

This document presents a novel approach to personalizing artwork recommendations on entertainment platforms like Netflix using Large Language Models (LLMs). Recognizing that users have diverse preferences and a single artwork may not appeal to everyone, the research addresses the challenge of user heterogeneity in visual content. The core idea involves post-training pre-existing LLMs to select the most preferred visual representation for a given title for each individual user, aiming to enhance user satisfaction and engagement. The methodology extends current text-based LLM recommendation systems to predict personalized user preferences among multiple visual options for a title. Experiments were conducted using Llama 3.1-8B models, which were post-trained with techniques like supervised fine-tuning (SFT) and Direct Preference Optimization (DPO), leveraging reasoning distillation from larger models. The results demonstrate significant improvements, achieving 3-5% better performance compared to the existing Netflix production model on a held-out dataset. This study highlights a promising direction for fine-grained content personalization, including artwork, synopses, and trailers, by effectively leveraging the advanced capabilities of LLMs to cater to varied user interests. The work distinguishes itself by uniquely focusing on personalized artwork selection, a sub-problem not extensively covered by prior LLM recommendation research. #LLM #Personalization #Netflix #ArtworkRecommendation #MachineLearning #DeepLearning #FineTuning #UserExperience #AI Donats:   / luxak   paper - https://arxiv.org/abs/2601.02764 subscribe - https://t.me/arxivpaper created with NotebookLM