Leveraging Large Language Models in Conversational Recommender Systems

TL;DR

This paper argues for a scalable, end-to-end Conversational Recommender System powered by Large Language Models, detailing RecLLM: a unified dialogue manager, tractable retrieval over massive item corpora, a joint ranking with natural-language explanations, interpretable natural-language user profiles, and a controllable LLM-based user simulator for synthetic data generation. It compares unified LLM control with modular designs, and it presents multiple retrieval strategies (Generalized Dual Encoder, Direct LLM Search, Concept-Based Search, and Search API Lookup) alongside a ranking/explanation module that produces rationales. A key contribution is the integration of persistent user profiles and the ability to simulate conversations to tune system modules at scale, enabling data-efficient learning in the absence of production logs. The RecLLM prototype on YouTube using LaMDA demonstrates fluent dialogue, diverse capabilities, and a concrete path toward practical deployment, while outlining future work on safety, debiasing, and broader multimodal support.

Abstract

A Conversational Recommender System (CRS) offers increased transparency and control to users by enabling them to engage with the system through a real-time multi-turn dialogue. Recently, Large Language Models (LLMs) have exhibited an unprecedented ability to converse naturally and incorporate world knowledge and common-sense reasoning into language understanding, unlocking the potential of this paradigm. However, effectively leveraging LLMs within a CRS introduces new technical challenges, including properly understanding and controlling a complex conversation and retrieving from external sources of information. These issues are exacerbated by a large, evolving item corpus and a lack of conversational data for training. In this paper, we provide a roadmap for building an end-to-end large-scale CRS using LLMs. In particular, we propose new implementations for user preference understanding, flexible dialogue management and explainable recommendations as part of an integrated architecture powered by LLMs. For improved personalization, we describe how an LLM can consume interpretable natural language user profiles and use them to modulate session-level context. To overcome conversational data limitations in the absence of an existing production CRS, we propose techniques for building a controllable LLM-based user simulator to generate synthetic conversations. As a proof of concept we introduce RecLLM, a large-scale CRS for YouTube videos built on LaMDA, and demonstrate its fluency and diverse functionality through some illustrative example conversations.

Figures (24)

• Figure 1: Overview of key contributions from RecLLM. (1) A dialogue management module uses an LLM to converse with the user, track context and make system calls such as submitting a request to a recommendation engine all as a unified language modeling task. (2) Various solutions are presented for tractable retrieval over a large item corpus within an LLM-based CRS. (3) A ranker module uses an LLM to match preferences extracted from the context of the conversation to item metadata and generate a slate of recommendations that is displayed to the user. The LLM also jointly generates explanations for its decisions that can be surfaced to the user. (4) Interpretable natural language user profiles are consumed by system LLMs to modulate session-level context and increase personalization. (5) A controllable LLM-based user simulator can be plugged into the CRS to generate synthetic conversations for tuning system modules.
• Figure 2: Screenshot of an LLM-based user simulator talking with RecLLM.
• Figure 3: RecLLM possesses many conversational capabilities such as the ability to retain context throughout a session, handle topic shifts and reference items from recommendation slates.
• Figure 4: A unified LLM dialogue management module. An LLM takes as input the full session context and outputs a sequence of messages ending in a terminal output that triggers a system action, such as a response to the user.
• Figure 5: Overview of large-scale retrieval in an LLM-based CRS.