EAGER-LLM: Enhancing LLMs as Recommenders through Exogenous Behavior-Semantic Integration (arXiv 2502.14735)

Forum topic · 小凯 · 2026-07-05

Summary

EAGER-LLM is a February 2025 arXiv paper (arXiv:2502.14735) by Minjie Hong, Yan Xia, Zehan Wang, Jieming Zhu, Ye Wang, Sihang Cai and colleagues that proposes a retriever–generator framework for building recommenders on top of large language models. The core idea is an exogenous integration of behavioral collaborative signals with the semantic space of the LLM, so that user-item interaction patterns are injected without requiring invasive end-to-end training of the language model itself. The framework demarcates collaborative signals from semantic alignment and uses them complementarily, aiming to improve recommendation accuracy while preserving the LLM's general semantic knowledge and keeping training and inference costs manageable. Positioned in the intersection of large-scale recommendation systems and LLM-based generative retrieval, the work addresses long-standing challenges in efficiency, scalability, and user intent understanding, and discusses interfaces with emerging paradigms such as tool calling and agentic pipelines. Open problems highlighted in this research area include evaluation credibility, latency and cost constraints, hallucination risks, and cross-lingual or multimodal extension. Readers should consult the original PDF for exact experimental numbers and dataset details.

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

Paper: arXiv:2502.14735
Date: February 2025
Authors: Minjie Hong, Yan Xia, Zehan Wang, Jieming Zhu, Ye Wang, Sihang Cai, et al. (11 authors total)
Category: Recommender Engines / LLM-based Recommendation

Key points

Proposes EAGER-LLM, a retriever–generator framework that adapts large language models for recommendation tasks.
Introduces exogenous integration: behavioral collaborative signals are merged with the LLM's semantic space externally, rather than through invasive end-to-end fine-tuning of the model.
Demarcates collaborative signals from semantic alignment and leverages them in a complementary manner.
Targets the classic tension in LLM-based recommenders between recommendation accuracy and preserving general semantic knowledge, while keeping training and inference costs tractable.
Situated within the broader shift from cascaded retrieve-rank-generate pipelines toward LLM-centric generative recommendation (Gen-Rec).

Context

Large-scale search and recommendation systems have long faced challenges in efficiency, scalability, and user intent understanding. Traditional pipelines separate retrieval, ranking, and generation, which is at odds with the natural-language interaction, multi-hop reasoning, and real-time knowledge expectations of the LLM era. EAGER-LLM addresses how to redraw the responsibility boundaries among retrieval, ranking, generation, and tool invocation in this new setting.

Related entries

360Brew: A Decoder-only Foundation Model for Personalized Ranking
Bridging Language and Items for Retrieval and Recommendation
Data-efficient Fine-tuning for LLM-based Recommendation (SIGIR 2024, DOI: 10.1145/3626772.3657807)

Takeaways for practitioners

1. Researchers: Reproduce the core comparisons and check whether statistical significance and compute cost are reported. 2. Engineers: Evaluate how pluggable components (encoder, reranker, planner) integrate with existing recommendation stacks; watch p99 latency budgets and embedding version management. 3. Product: Focus on user-perceivable gains (latency, answer credibility, multi-turn consistency) rather than offline metrics alone.

> Note: This post is an index-level summary. Consult the original PDF for exact experimental results and dataset details.

References

Original paper: EAGER-LLM: Enhancing Large Language Models as Recommenders through Exogenous Behavior-Semantic Integration. https://arxiv.org/abs/2502.14735