Bidirectional Knowledge Distillation for Enhancing Sequential Recommendation with Large Language Models

TL;DR

LLMD4Rec introduces a bidirectional mutual distillation framework that enables dynamic collaboration between LLM-centric and CRM-based sequential recommender systems. By alternating training and aligning output distributions through temperature-scaled KL divergences and sample-wise weighting, it achieves mutual enhancement without adding trainable parameters or increasing inference costs. The approach injects CRM signals into the LLM (information augmentation) and distills LLM semantics into the CRM (output distribution distillation), forming a virtuous loop that improves both components. Empirical results on real-world datasets show substantial gains over strong baselines, with efficient training and deployment suitable for scalable, real-time applications.

Abstract

Large language models (LLMs) have demonstrated exceptional performance in understanding and generating semantic patterns, making them promising candidates for sequential recommendation tasks. However, when combined with conventional recommendation models (CRMs), LLMs often face challenges related to high inference costs and static knowledge transfer methods. In this paper, we propose a novel mutual distillation framework, LLMD4Rec, that fosters dynamic and bidirectional knowledge exchange between LLM-centric and CRM-based recommendation systems. Unlike traditional unidirectional distillation methods, LLMD4Rec enables iterative optimization by alternately refining both models, enhancing the semantic understanding of CRMs and enriching LLMs with collaborative signals from user-item interactions. By leveraging sample-wise adaptive weighting and aligning output distributions, our approach eliminates the need for additional parameters while ensuring effective knowledge transfer. Extensive experiments on real-world datasets demonstrate that LLMD4Rec significantly improves recommendation accuracy across multiple benchmarks without increasing inference costs. This method provides a scalable and efficient solution for combining the strengths of both LLMs and CRMs in sequential recommendation systems.

Figures (4)

• Figure 1: The illustration of the structural comparison between and LLM-centric RM and CRM distillation, along with our LLMD4Rec's process of mutual distillation training using both components. In (a), the pre-trained CRM supplies item ID embeddings, while the LLM-centric RM is optimized by fine-tuning its LoRA modules and projection layers. In (b), the pre-finetuned LLM-centric RM provides representations from its embedding or top-K ranking result to guide the training and optimization of the CRM. In (c), the top and bottom information from the CRM is utilized to guide the fine-tuning of the LLM, while the more accurate recommendation scores generated by the LLM-centric RM further contribute to optimizing the recommendation performance of the CRM.
• Figure 2: Overview of the proposed LLMD4Rec framework, depicting the directional flow of information and the iterative update mechanism of the LLM-centric RM. The update mechanism for the CRM is similar.
• Figure 3: Ablation study of the loop mechanism, distillation sample weight, and ID embedding update in LLMD4Rec.
• Figure 4: Overview of the proposed LLMD4Rec framework