Technical Summary: NaviGen – Navigating User Behavior toward Personalized Multimodal Generation

1. Problem Definition

The paper addresses a critical gap in modern Artificial Intelligence Generated Content (AIGC) pipelines. While current text-to-image and text-to-video generators produce high-fidelity outputs, they rely on well-formed textual instructions. End users, however, rarely articulate specific visual details; their preferences are implicit in their interaction histories (clicks, views, purchases).

This creates a "Preference-to-Instruction Gap." Existing methods either assume a perfect prompt exists or rely on generic prompt expansion, failing to translate a user's specific, implicit behavioral history into a concrete, generation-ready instruction. The authors identify two core challenges:

1. Representation Gap (C1): How to encode user behavior into a form legible to Large Language Models (LLMs) that preserves behavioral structure without being verbose or semantically opaque.
2. Capability Gap (C2): How to equip models with the skill to write specific creation instructions based on preference reasoning, a capability not naturally acquired through standard pretraining or behavior data.

2. Methodology: NaviGen

The authors propose NaviGen, a unified framework that converts user interaction sequences into executable AIGC instructions. The architecture consists of three main components:

A. Dual-Identifier Behavior Encoding

To bridge the representation gap, NaviGen encodes each item in a user's history using a dual-identifier scheme within a single token stream:

• Collaborative Identifier (CID): A three-level residual vector quantization token sequence derived from the item's metadata. It captures the item's role in the interaction graph and collaborative patterns, serving as a compact behavioral substrate.
• Textual Identifier (TID): A variable-length (up to 10 terms), ordered list of concise semantic phrases (e.g., subject, scene, attributes) compressed from the original caption. This acts as a controllable semantic bridge.
• Coupling: The CID and TID are coupled to provide both behavioral signal and semantic grounding, allowing the model to reason over preferences without being overwhelmed by raw text.

B. Two-Stage Training Pipeline

NaviGen employs a Supervised Fine-Tuning (SFT) + Reinforcement Learning (RL) pipeline to close the capability gap.

Stage 1: Reasoning-Infused SFT

• Cold-Start Initialization: CID embeddings are initialized by freezing the backbone and training only on auxiliary tasks (CID2TID, TID2CID, CID2CID) to ensure stable representation acquisition.
• Reasoning-Augmented Full Finetuning: The model is trained on a synthesized dataset where instructions are generated via an evolutionary search process. An LLM judge selects and evolves candidate instructions (via crossover and mutation) to align with user semantics.
• Chain-of-Thought (CoT): The training data includes reasoning traces that explain how user preferences evolve along the history (e.g., "User moved from humorous elf interactions to romantic conflict"), teaching the model to infer trends rather than just paraphrase history.

Stage 2: Multi-Task Reinforcement Learning (GRPO)
The model is further optimized using Group Relative Policy Optimization (GRPO) with two complementary reward signals:

1. Hierarchical CID Reward: Evaluates the correctness of the predicted next-item CID at three levels of granularity (coarse to fine), prioritizing family-level accuracy over exact variant precision.
2. Instruction-Aware Reward: A triangular self-consistency check involving:
   • Quality: Specificity, creativity, and visual generatability (evaluated by an LLM judge).
   • Alignment: Ensuring the generated instruction anchors to the target semantics, remains self-consistent with its own prediction, and aligns with ground truth.

3. Key Contributions

1. NaviGen Framework: A unified pipeline that bridges behavioral modeling and controllable content generation, transforming implicit user history into generation-ready instructions.
2. Dual-Identifier Representation: A novel scheme coupling residual-quantized CIDs (behavioral substrate) with ordered, length-flexible TIDs (semantic bridge) within a single token stream.
3. SFT+RL Training Strategy: A method requiring no human-written instructions, utilizing evolutionary search for supervision synthesis and GRPO with hierarchical and triangular rewards to enforce closed-loop self-consistency.
4. Empirical Validation: Demonstrated improvements across product, game, and short-video domains in personalized image/video generation quality, next-item prediction accuracy, and instruction specificity.

4. Experimental Results

The authors evaluated NaviGen on three datasets: Product (Amazon reviews), Games (Amazon reviews), and Short Videos (OpenOneRec benchmark).

• Personalized Generation Performance: NaviGen achieved state-of-the-art results in most non-oracle comparisons.
  • Images: It led in Novelty, Aesthetic, and Relevance across all domains, with particularly strong performance in Games and Short Videos.
  • Videos: It achieved the best results in 9 out of 12 non-oracle comparisons, balancing frame-level consistency with preference-specific relevance.
  • Comparison to Oracle: While an "Oracle" (conditioned on ground-truth target semantics) achieved the highest Relevance, NaviGen surpassed it in Aesthetic scores for Product and Short Video images, suggesting that target semantics alone do not guarantee superior perceptual quality.
• Collaborative Identifier Prediction: NaviGen outperformed strong baselines (SASRec, TIGER, LC-Rec, OneRec) in predicting the next item's CID, particularly in Recall@20 and NDCG@20 for Product and Short Video domains.
• Ablation Studies: Removing TID grounding, CID initialization, reasoning supervision, or GRPO optimization consistently degraded performance, confirming the necessity of each component for both generation quality and behavioral modeling.

5. Significance and Claims

The paper claims that NaviGen establishes a unified path from implicit user behavior to controllable multimodal synthesis. Its primary significance lies in:

• Solving the Input Bottleneck: It addresses the fundamental limitation that end users cannot write detailed prompts, by automating the translation of behavioral signals into high-quality instructions.
• Decoupling Behavior and Semantics: By separating behavioral coding (CID) from semantic grounding (TID), the model treats preference prediction as an intermediate reasoning step rather than an end goal, allowing for more robust generalization.
• Enhancing AIGC Utility: The framework enables downstream generators to produce content that is not only high-fidelity but also specifically aligned with individual user tastes, moving beyond generic generation to true personalization.

The authors maintain a modest stance, noting that the work is an "assistive layer for creative control" and emphasizing that practical deployment requires explicit user consent and privacy safeguards regarding the use of interaction histories.