NaviGen Personalizes Multimodal Generation from User Behavior

Modern AI-generated content (AIGC) pipelines excel at producing high-fidelity images and videos, but they typically require precise creation instructions. End-users, however, often struggle to articulate detailed visual preferences, leading to a misalignment between generator output and user demand. This research introduces NaviGen, a system designed for personalized content generation that translates a user's interaction history into actionable instructions for downstream synthesis. NaviGen addresses two key challenges: encoding user behavior in a format understandable by language reasoning, and enabling the model to acquire instruction-writing skills not present in its pretraining or behavioral data. It achieves this by representing each item with a dual identifier, combining a collaborative code and a textual code into a single token stream, serving as both a behavioral substrate and a semantic bridge. The system employs a two-stage pipeline involving Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL). This process first distills preference reasoning and instruction writing from evolutionarily searched supervision, then aligns the generation with user intent through hierarchical and self-consistent rewards. Experiments across product, game, and short-video domains demonstrate that NaviGen significantly improves personalized image and video generation, enhances next-item prediction, and produces more specific, relevant, and visually generatable instructions.