MVIGER: Multi-View Variational Integration of Complementary Knowledge for Generative Recommender

Tongyoung Kim, Soojin Yoon, SeongKu Kang, Jinyoung Yeo, Dong Ha Lee

Abstract: Language Models (LMs) have been widely used in recommender systems to incorporate textual information of items into item IDs, leveraging their advanced language understanding and generation capabilities. Recently, generative recommender systems have utilized the reasoning abilities of LMs to directly generate index tokens for potential items of interest based on the user's interaction history. To inject diverse item knowledge into LMs, prompt templates with detailed task descriptions and various indexing techniques derived from diverse item information have been explored. This paper focuses on the inconsistency in outputs generated by variations in input prompt templates and item index types, even with the same user's interaction history. Our in-depth quantitative analysis reveals that preference knowledge learned from diverse prompt templates and heterogeneous indices differs significantly, indicating a high potential for complementarity. To fully exploit this complementarity and provide consistent performance under varying prompts and item indices, we propose MVIGER, a unified variational framework that models selection among these information sources as a categorical latent variable with a learnable prior. During inference, this prior enables the model to adaptively select the most relevant source or aggregate predictions across multiple sources, thereby ensuring high-quality recommendation across diverse template-index combinations. We validate the effectiveness of MVIGER on three real-world datasets, demonstrating its superior performance over existing generative recommender baselines through the effective integration of complementary knowledge.

Introduction

Generative recommender systems represent a paradigm shift in personalized discovery. Unlike traditional discriminative models that rank a candidate set, generative models use Language Models (LMs) to "write out" the identifiers of recommended items. However, these systems are notoriously sensitive to input variations. A slight change in the wording of a prompt or the method used to index items (e.g., using semantic descriptions vs. collaborative behavior) can lead to wildly inconsistent results.

The authors of MVIGER (Multi-View Variational Integration of GEnerative Recommender) argue that these inconsistencies are not just noise but represent different "views" of user preferences. By treating prompt-index combinations as complementary sources of knowledge, MVIGER uses a variational framework to adaptively merge these perspectives, resulting in more robust and accurate recommendations.

Methodology: A Three-Step Pipeline

MVIGER introduces a structured approach to unify diverse recommendation perspectives:

  1. Heterogeneous Indexing: Items are assigned multiple "names" or indices using a Residual Quantized Variational Autoencoder (RQ-VAE). Specifically, it generates CeID (Collaborative IDs) based on user-item interaction patterns and SeID (Semantic IDs) based on item metadata like titles and descriptions.
  2. Multi-View Variational Training: The model treats the choice of a specific prompt-index combination as a categorical latent variable ($z$). By optimizing the Evidence Lower Bound (ELBO), the LM learns to associate specific user histories with the most informative "views."
  3. Flexible Inference: At test time, a "Prior Network" estimates the probability of each view. The system can perform Single View Inference (selecting the most probable view for speed) or Multi View Inference (aggregating predictions across views for higher accuracy).
The MVIGER framework: from multi-view item index generation to variational training and inference.

Experimental Setup

The model was rigorously tested on three standard benchmarks: Amazon Beauty, Amazon Sports, and Yelp. These datasets were processed using standard "5-core" filtering to ensure data quality. The evaluation followed a leave-one-out strategy, utilizing metrics such as Hit Ratio (H@K) and Normalized Discounted Cumulative Gain (N@K) to assess recommendation quality.

Results and Key Insights

The core finding of the study is the high degree of complementarity between different views. Quantitative analysis showed that different indexing strategies often correctly predict different sets of items, with less than 50% overlap in many cases.

Main experimental results comparing MVIGER against state-of-the-art generative and discriminative baselines.

As shown in the performance table, MVIGER consistently outperforms state-of-the-art generative baselines like TIGER, P5, and EAGER. For example, on the Amazon Beauty dataset, MVIGER achieved a Hit@10 of 0.0878, significantly better than the previous best of 0.0836.

Key takeaways from the results include:

  • Integration over Scale: MVIGER, built on the T5-small architecture, often outperformed much larger models like EAGER-LLM (based on Llama-7b), proving that intelligent knowledge integration can be more effective than simply increasing parameter count.
  • Robustness: The variational framework effectively mitigates the "consistency crisis," providing stable performance even when prompt templates are varied.
  • Efficiency: The Prior Network adds negligible overhead (approx. 0.01s per user), making the multi-view approach viable for real-world production environments.

Conclusion

MVIGER successfully addresses the sensitivity of generative recommenders by embracing the diversity of input prompts and item indices. By modeling these variations as complementary views within a variational framework, it provides a robust and high-performing solution for modern recommendation tasks. Future research will look into scaling the number of views and applying this framework to cold-start recommendation scenarios.

Summary updated

Abstract

Language Models (LMs) have been widely used in recommender systems to incorporate textual information of items into item IDs, leveraging their advanced language understanding and generation capabilities. Recently, generative recommender systems have utilized the reasoning abilities of LMs to directly generate index tokens for potential items of interest based on the user's interaction history. To inject diverse item knowledge into LMs, prompt templates with detailed task descriptions and various indexing techniques derived from diverse item information have been explored. This paper focuses on the inconsistency in outputs generated by variations in input prompt templates and item index types, even with the same user's interaction history. Our in-depth quantitative analysis reveals that preference knowledge learned from diverse prompt templates and heterogeneous indices differs significantly, indicating a high potential for complementarity. To fully exploit this complementarity and provide consistent performance under varying prompts and item indices, we propose MVIGER, a unified variational framework that models selection among these information sources as a categorical latent variable with a learnable prior. During inference, this prior enables the model to adaptively select the most relevant source or aggregate predictions across multiple sources, thereby ensuring high-quality recommendation across diverse template-index combinations. We validate the effectiveness of MVIGER on three real-world datasets, demonstrating its superior performance over existing generative recommender baselines through the effective integration of complementary knowledge.
BibTeX
@inproceedings{kim2026mviger,
  author = {Tongyoung Kim and Soojin Yoon and SeongKu Kang and Jinyoung Yeo and Dong Ha Lee},
  title = {{MVIGER: Multi-View Variational Integration of Complementary Knowledge for Generative Recommender}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2026},
  doi = {10.1145/3805712.3809591},
}
Figure