CARE: Contextual Adaptation of Recommenders for LLM-based Conversational Recommendation

2025-08-19

AI Summary

Existing LLM-based conversational recommender systems (CRS) face two key challenges: (1) difficulty in constraining recommendations to the target item space under zero- or few-shot settings, leading to low accuracy; and (2) neglect of collaborative relationships among entities, exacerbating popularity bias. To address these, we propose CARE, the first framework enabling synergistic reasoning between an LLM and an external recommendation system (RS). CARE leverages the external RS to supply domain-aware candidate items and entity-relation priors, then performs context-aware, entity-level joint re-ranking using conversational history. This design jointly mitigates item-space mismatch and suppresses popularity bias. Evaluated on ReDial and INSPIRED, CARE achieves 54% and 25% improvements in recommendation accuracy over state-of-the-art zero-/few-shot methods, respectively.

Technology Category

• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Knowledge Representation and Reasoning: Common-Sense Reasoning

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

We tackle the challenge of integrating large language models (LLMs) with external recommender systems to enhance domain expertise in conversational recommendation (CRS). Current LLM-based CRS approaches primarily rely on zero- or few-shot methods for generating item recommendations based on user queries, but this method faces two significant challenges: (1) without domain-specific adaptation, LLMs frequently recommend items not in the target item space, resulting in low recommendation accuracy; and (2) LLMs largely rely on dialogue context for content-based recommendations, neglecting the collaborative relationships among entities or item sequences. To address these limitations, we introduce the CARE (Contextual Adaptation of Recommenders) framework. CARE customizes LLMs for CRS tasks, and synergizes them with external recommendation systems. CARE (a) integrates external recommender systems as domain experts, producing recommendations through entity-level insights, and (b) enhances those recommendations by leveraging contextual information for more accurate and unbiased final recommendations using LLMs. Our results demonstrate that incorporating external recommender systems with entity-level information significantly enhances recommendation accuracy of LLM-based CRS by an average of 54% and 25% for ReDial and INSPIRED datasets. The most effective strategy in the CARE framework involves LLMs selecting and reranking candidate items that external recommenders provide based on contextual insights. Our analysis indicates that the CARE framework effectively addresses the identified challenges and mitigates the popularity bias in the external recommender.

Problem

• Integrating LLMs with external recommender systems for conversational recommendation
• Addressing low accuracy from LLMs recommending non-existent domain items
• Mitigating bias by combining entity-level insights with contextual adaptation

Innovation

• Integrates external recommender systems with entity-level insights
• Enhances recommendations using contextual information via LLMs
• Selects and reranks candidate items based on context