IntRR: A Framework for Integrating SID Redistribution and Length Reduction

Zesheng Wang, Longfei Xu, Weidong Deng, Huimin Yan, Kaikui Liu, Xiangxiang ChuarXiv 2026 · DOI

TL;DR

This paper introduces IntRR, a novel framework that tackles the inefficiencies in generative recommendation systems by dynamically redistributing semantic identifiers (SIDs) and reducing sequence lengths. IntRR achieves significant improvements in recommendation accuracy and drastically cuts down computational costs and inference latency, making real-time recommendations more feasible.

What Problem Does This Paper Address?

Traditional generative recommendation (GR) models reframe recommendation as a sequence-to-item generation task, often using discrete Semantic IDs (SIDs) to represent items. However, current methods face two major challenges:

  1. Suboptimal SID Generation: SIDs are typically generated offline based on item content, ignoring crucial collaborative user interaction patterns. This "static bottleneck" limits the model's ability to adapt item representations to user preferences.
  2. Sequence Length Inflation: Hierarchical SIDs are often flattened into long token sequences. This leads to a quadratic increase in computational complexity for Transformer-based models, resulting in prohibitive training and inference costs.

These issues create representation ceilings and hinder the scalability of GR systems for real-world applications.

Methods

IntRR proposes a framework that addresses these problems by integrating two key mechanisms:

  • Objective-Aligned SID Redistribution: Instead of relying on static SIDs, IntRR uses item-specific Unique IDs (UIDs) as collaborative anchors. A novel Recursive-Assignment Network (RAN) dynamically redistributes semantic weights across hierarchical codebook layers. This allows the model to refine coarse semantic priors from the offline stage with collaborative signals, creating more expressive and dynamically optimized item representations.
  • Structural Length Reduction: The RAN handles the SID hierarchy recursively, avoiding the need to flatten sequences. This means each item is represented by a single token in the backbone, drastically reducing the input sequence length. The RAN internalizes the hierarchical navigation, offloading this complexity from the main recommendation backbone.

The framework also employs a joint optimization strategy that balances semantic alignment (ensuring SIDs remain structurally coherent) and recommendation accuracy.

Datasets

The paper evaluates IntRR on three widely-used Amazon benchmarks:

  • Beauty
  • Sports
  • Toys

These datasets are processed using a standard 5-core filter protocol and a leave-one-out evaluation strategy.

Key Insights and Ideas

  • Dynamic SID Refinement is Crucial: The static nature of SIDs generated offline is a significant limitation. IntRR demonstrates that allowing the recommendation model to dynamically adjust SID weights based on collaborative signals leads to better representations.
  • Recursion as a Solution for Hierarchy: Instead of flattening hierarchical structures, a recursive approach within the RAN effectively models the SID hierarchy without increasing sequence length. This is a key insight for computational efficiency.
  • UIDs as Collaborative Anchors: Leveraging UIDs as a bridge between collaborative signals and semantic identifiers allows for a more nuanced and adaptive representation learning process.
  • Decoupling for Efficiency: By offloading the complex SID navigation to the RAN, the main recommendation backbone can focus on learning user-item interactions from a compact sequence, leading to substantial speedups.

Key Results

  • Improved Accuracy: IntRR consistently outperforms baseline models across various indexing strategies (RK-Means, VQ-VAE, RQ-VAE) and backbone architectures (Transformer, HSTU), achieving significant gains in recommendation accuracy (e.g., up to 63.1% improvement on Amazon Toys with HSTU and RQ-VAE).
  • Drastic Efficiency Gains:
    • Training Throughput: IntRR significantly boosts throughput, with improvements of up to 75.0% for the Transformer backbone.
  • VRAM Reduction: Peak VRAM usage is reduced by up to 68.2% for the Transformer backbone due to the shorter sequence length.
  • Inference Latency: IntRR achieves substantial speedups in inference latency, growing from 1.27x to 2.93x as beam width increases, making real-time recommendations more feasible.

Limitations and Future Work

  • Sensitivity to Weighting: The performance of IntRR is sensitive to the weight (λ) balancing semantic anchoring and redistribution flexibility. An optimal balance (around λ=0.1) is crucial; too much freedom leads to "semantic drift," while too much constraint makes representations rigid.
  • Future Work: The paper suggests that IntRR's versatility makes it a strong candidate for integration into various existing generative recommendation pipelines. Further exploration of its application in more complex scenarios or with different types of hierarchical identifiers could be beneficial.

Summary updated

Abstract

Generative Recommendation (GR) has emerged as a transformative paradigm that reformulates the traditional cascade ranking system into a sequence-to-item generation task, facilitated by the use of discrete Semantic IDs (SIDs). However, current SIDs are suboptimal as the indexing objectives (Stage 1) are misaligned with the actual recommendation goals (Stage 2). Since these identifiers remain static (Stage 2), the backbone model lacks the flexibility to adapt them to the evolving complexities of user interactions. Furthermore, the prevailing strategy of flattening hierarchical SIDs into token sequences leads to sequence length inflation, resulting in prohibitive computational overhead and inference latency. To address these challenges, we propose IntRR, a novel framework that integrates objective-aligned SID Redistribution and structural Length Reduction. By leveraging item-specific Unique IDs (UIDs) as collaborative anchors, this approach dynamically redistributes semantic weights across hierarchical codebook layers. Concurrently, IntRR handles the SID hierarchy recursively, eliminating the need to flatten sequences. This ensures a fixed cost of one token per item. Extensive experiments on benchmark datasets demonstrate that IntRR yields substantial improvements over representative generative baselines, achieving superior performance in both recommendation accuracy and efficiency.
BibTeX
@misc{wang2026intrr,
  author = {Zesheng Wang and Longfei Xu and Weidong Deng and Huimin Yan and Kaikui Liu and Xiangxiang Chu},
  title = {{IntRR: A Framework for Integrating SID Redistribution and Length Reduction}},
  year = {2026},
  eprint = {2602.20704},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.20704},
}
Figure