IntRR: A Framework for Integrating SID Redistribution and Length Reduction

Zesheng Wang¹, Longfei Xu¹†, Weidong Deng, Huimin Yan, Kaikui Liu, Xiangxiang Chu

AMAP, Alibaba Group

¹Equal contribution.         †Corresponding author and project lead.

📖 Overview

We introduce IntRR, a novel generative recommendation (GR) framework designed to break the representation ceiling and computational bottlenecks of current Semantic ID (SID)-based systems. Within a two-stage paradigm (Semantic Indexing &Generative Learning), IntRR optimizes Stage 2 by internalizing the hierarchical and flattened SIDs into the backbone, achieving deep collaborative-semantic integration while maintaining a constant one-token-per-item complexity.

💡 IntRR: Core Framework

The core of IntRR is the Recursive-Assignment Network (RAN), which functions as a differentiable bridge between collaborative signals and semantic structures through two key mechanisms:

• Adaptive SID Redistribution: Utilizes item's Unique IDs (UIDs) as collaborative signals to dynamically refine semantic weights. This mechanism aligns content-based identifiers from Stage 1 with recommendation goals, breaking the "static ceiling" of traditional SIDs.
• Structural Length Reduction: Internalizes the item's hierarchical navigation of SIDs within a recursive path. It reduces backbone's sequence length to a single token per item, eliminating multi-step inference bottleneck and significantly enhancing system throughput.

📊 Performance & Efficiency

IntRR yields substantial improvements in both recommendation accuracy and system scalability across multiple benchmarks.

1. Recommendation Accuracy

Overall performance comparison across diverse indexing methods (RK-Means, VQ-VAE, RQ-VAE) and backbones (Transformer, HSTU). IntRR consistently achieves superior recommendation accuracy and outperforms representative generative baselines.

2. Efficiency & Complexity

Efficiency comparison in terms of training throughput, memory consumption, and inference latency. By bypassing SID flattening and the multi-pass inference bottleneck, IntRR delivers significant gains in system scalability.

✨ Visualizing SID Redistribution

Our analysis demonstrates that RAN adaptively steers item representations. Even for items sharing identical initial SIDs, IntRR triggers semantic weight redistribution based on collaborative interaction patterns, yielding more refined and unique item embeddings.

📂 Repository Structure

IntRR/
├── configs/                 # Configuration files
│   ├── callbacks/          # PyTorch Lightning callbacks
│   ├── experiment/         # Experiment configurations (training/inference)
│   ├── extras/             # Extra configurations
│   ├── logger/             # Logging configurations
│   ├── paths/              # Path configurations
│   └── trainer/            # Trainer configurations
├── refs/                   # Reference images and figures
├── src/                    # Source code
│   ├── components/         # Core components
│   │   ├── clustering_initializers.py
│   │   ├── distance_functions.py
│   │   ├── eval_metrics.py
│   │   ├── loss_functions.py
│   │   ├── optimizer.py
│   │   ├── quantization_strategies.py
│   │   ├── scheduler.py
│   │   └── training_loop_functions.py
│   ├── models/             # Model implementations
│   │   ├── components/     # Model components
│   │   └── modules/        # Model modules
│   ├── modules/            # Neural network modules
│   │   └── clustering/     # Clustering algorithms
│   └── utils/              # Utility functions
├── gen_sid.sh              # Script to generate Semantic IDs
├── run_intrr.sh            # Script to run IntRR training
├── run_tiger.sh            # Script to run TIGER baseline
├── requirements.txt        # Python dependencies
└── README.md               # This file

📦 Installation

Prerequisites

• Python 3.10+

• CUDA-compatible GPU (recommended)

🎯 Quick Start

Prerequisites

For environment setup and data preparation, please refer to the GRID (Generative Recommendation with Semantic IDs) repository.

1. Generate Semantic IDs

Generate Semantic IDs and Update dataset_config.sh

sh  gen_sid.sh  --datasets  sports  --sid-methods  rkmeans

2. Train IntRR with Semantic IDs

Train the recommendation model using the learned semantic IDs:

sh  run_intrr.sh  --datasets  sports  --seeds  42  --sid-type  rkmeans

🙏 Acknowledgments

This work builds upon the GRID framework by Snap Research. We thank the GRID team for their open-source contributions to the generative recommendation community, which provides a solid foundation for SID-based generative recommendation research.

📚 Citation

If you find our paper and code helpful for your research, please consider starring our repository ⭐ and citing our work ✏️.

@misc{wang2026intrrframeworkintegratingsid,
      title={IntRR: A Framework for Integrating SID Redistribution and Length Reduction}, 
      author={Zesheng Wang and Longfei Xu and Weidong Deng and Huimin Yan and Kaikui Liu and Xiangxiang Chu},
      year={2026},
      eprint={2602.20704},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2602.20704}, 
}