NaviGen Personalizes Multimodal Generation Using User Behavior.

Modern AI-generated content (AIGC) systems excel at producing high-quality images and videos, but they typically require precise input instructions. End-users, however, often struggle to articulate detailed visual preferences, leading to a mismatch between their desires and the generated output. This paper explores personalized content generation, aiming to convert a user's past interactions into clear, actionable instructions for synthesis. The NaviGen framework tackles two primary hurdles: effectively encoding user behavior in a format understandable by language models, and enabling models to generate these instructions, a skill not inherent in pre-training or standard behavioral data. NaviGen represents each interaction item with a dual identifier, combining collaborative and textual codes into a single token stream, serving as both a behavioral substrate and a semantic bridge. A two-stage pipeline, involving supervised fine-tuning (SFT) followed by reinforcement learning (RL), is then employed. This pipeline first distills preference reasoning and instruction writing from carefully evolved supervision, and subsequently aligns the generation process with explicit user intent using hierarchical and self-consistent rewards. Evaluations across product, game, and short-video domains demonstrate that NaviGen enhances personalized image and video generation, improves next-item prediction accuracy, and produces instructions that are more specific, relevant, and visually feasible.