EAGER-LLM: Enhancing Large Language Models as Recommenders through Exogenous Behavior-Semantic Integration

TL;DR

This paper tackles the mismatch between LLM linguistic semantics and collaborative semantics in recommender systems by proposing EAGER-LLM, a decoder-only framework that non-intrusively integrates endogenous and exogenous behavioral and semantic signals. It introduces Dual-source Knowledge-rich Item Indices (DKI) based on hierarchical K-Means to compress vast candidate sets into a few tokens, and Non-Invasive Multiscale Alignment Reconstruction Tasks (Global Contrast Decompression Task and Comprehensive Interaction Modeling Task) guided by Decompression Guidance Projectors, complemented by a multi-stage Annealing Adapter to balance recommendation accuracy with comprehension. The approach is validated on three public benchmarks, demonstrating state-of-the-art performance and clear ablations showing the importance of each component (DKI, GCT, AAT). The work advances practical LLM-based recommendations by enabling richer exogenous-signal integration with scalable item representations, achieving robust improvements across diverse domains.

Abstract

Large language models (LLMs) are increasingly leveraged as foundational backbones in the development of advanced recommender systems, offering enhanced capabilities through their extensive knowledge and reasoning. Existing llm-based recommender systems (RSs) often face challenges due to the significant differences between the linguistic semantics of pre-trained LLMs and the collaborative semantics essential for RSs. These systems use pre-trained linguistic semantics but learn collaborative semantics from scratch via the llm-Backbone. However, LLMs are not designed for recommendations, leading to inefficient collaborative learning, weak result correlations, and poor integration of traditional RS features. To address these challenges, we propose EAGER-LLM, a decoder-only llm-based generative recommendation framework that integrates endogenous and exogenous behavioral and semantic information in a non-intrusive manner. Specifically, we propose 1)dual-source knowledge-rich item indices that integrates indexing sequences for exogenous signals, enabling efficient link-wide processing; 2)non-invasive multiscale alignment reconstruction tasks guide the model toward a deeper understanding of both collaborative and semantic signals; 3)an annealing adapter designed to finely balance the model's recommendation performance with its comprehension capabilities. We demonstrate EAGER-LLM's effectiveness through rigorous testing on three public benchmarks.

Figures (6)

• Figure 1: A framework of LLM-as-Recommender. It frames recommendation as next token generation tasks, which directly generate target items.
• Figure 2: The illustration of Dual-source Knowledge-rich Item Indice. Incorporate massive exogenous behavioral semantic signals into project indices with extremely high compression ratios.
• Figure 3: Overview of Non-Invasive Multiscale Alignment Reconstruction Tasks. Includes Global Contrast Decompression Task (GCT), Comprehensive Interaction Modeling Task (CIT). These tasks facilitate the LLM-Backbone’s comprehension of complex exogenous signals from these densely packed, knowledge-rich tokens in a non-invasive manner, and to integrate the llm's reasoning capabilities for deep understanding of recommendations.
• Figure 4: The illustration of multi-stage training & inference process.
• Figure 5: The performance of our architecture (w/o GCT, AAT), under indexing with different exogenous signal compositions.