TailorMind: Towards Preference-Aligned Multimodal Content Generation

Traditional personalized content systems are often limited by the availability of existing user-generated content (UGC), leading to delays in matching user interests with fresh material. This paper addresses the challenge of synthesizing user-tailored multimodal content on demand, specifically focusing on how to derive actionable, preference-aligned profiles from sparse behavioral data and maintain semantic consistency throughout the generation process.

The authors propose TailorMind, a framework that integrates collaborative preference modeling with controllable multimodal generation. The system operates in three main stages:

1. Hypergraph Collaborative Filtering: To overcome data sparsity, the model constructs a hypergraph connecting users, tags, and items, allowing for the propagation of collaborative signals to enrich user history.
2. Iterative Profile Optimization: Instead of relying on static prompts, TailorMind uses a textual-gradient descent mechanism. It treats the user profile as a mutable text string and iteratively updates it based on ranking-error feedback from actual user interactions.
3. Controlled Generation: The system employs retrieval-augmented style control, grounding outputs in authentic UGC patterns, and a cross-modal cohesion reflection mechanism that monitors and corrects semantic drift between generated text and visual modalities.

TailorMind was evaluated using TailorBench, a new benchmark constructed from Rednote, Bilibili, and Hupu data. The results demonstrate that TailorMind achieves superior performance in coherence, novelty, and aesthetic quality compared to representative generation baselines and ground-truth UGC. Notably, the framework achieved up to 29% gains in recall for reranking tasks, confirming that the optimized textual profiles effectively capture nuanced user preferences.