Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Selective LLM-Guided Regularization for Enhancing Recommendation Models

Shanglin Yang, Zhan Shi

TL;DR

The paper tackles the challenge of leveraging large language models (LLMs) in recommender systems under data sparsity, where standalone LLMs are expensive and global distillation can be harmful. It introduces Selective LLM-Guided Regularization (S-LLMR), which uses offline LLM scoring to generate soft rankings $s^{LLM}_{u,i}$ and a learnable gating mechanism to apply a pairwise ranking regularizer only in regions where LLM signals are reliable, via the objective $\\mathcal{L}=\mathcal{L}_{rec}+\lambda\\mathcal{L}_{LLM}$. Empirical results across six backbone architectures and three Amazon domains show consistent AUC improvements, with particularly strong gains in cold-start and long-tail scenarios, and ablation studies confirm the importance of reliability-aware gating and pairwise supervision over global imitation. The approach is model-agnostic and computation-efficient, achieving knowledge transfer without increasing inference cost and offering practical impact for robust recommendations in sparse or semantically challenging settings.

Abstract

Large language models provide rich semantic priors and strong reasoning capabilities, making them promising auxiliary signals for recommendation. However, prevailing approaches either deploy LLMs as standalone recommender or apply global knowledge distillation, both of which suffer from inherent drawbacks. Standalone LLM recommender are costly, biased, and unreliable across large regions of the user item space, while global distillation forces the downstream model to imitate LLM predictions even when such guidance is inaccurate. Meanwhile, recent studies show that LLMs excel particularly in re-ranking and challenging scenarios, rather than uniformly across all contexts.We introduce Selective LLM Guided Regularization, a model-agnostic and computation efficient framework that activates LLM based pairwise ranking supervision only when a trainable gating mechanism informing by user history length, item popularity, and model uncertainty predicts the LLM to be reliable. All LLM scoring is performed offline, transferring knowledge without increasing inference cost. Experiments across multiple datasets show that this selective strategy consistently improves overall accuracy and yields substantial gains in cold start and long tail regimes, outperforming global distillation baselines.

Selective LLM-Guided Regularization for Enhancing Recommendation Models

TL;DR

The paper tackles the challenge of leveraging large language models (LLMs) in recommender systems under data sparsity, where standalone LLMs are expensive and global distillation can be harmful. It introduces Selective LLM-Guided Regularization (S-LLMR), which uses offline LLM scoring to generate soft rankings and a learnable gating mechanism to apply a pairwise ranking regularizer only in regions where LLM signals are reliable, via the objective . Empirical results across six backbone architectures and three Amazon domains show consistent AUC improvements, with particularly strong gains in cold-start and long-tail scenarios, and ablation studies confirm the importance of reliability-aware gating and pairwise supervision over global imitation. The approach is model-agnostic and computation-efficient, achieving knowledge transfer without increasing inference cost and offering practical impact for robust recommendations in sparse or semantically challenging settings.

Abstract

Large language models provide rich semantic priors and strong reasoning capabilities, making them promising auxiliary signals for recommendation. However, prevailing approaches either deploy LLMs as standalone recommender or apply global knowledge distillation, both of which suffer from inherent drawbacks. Standalone LLM recommender are costly, biased, and unreliable across large regions of the user item space, while global distillation forces the downstream model to imitate LLM predictions even when such guidance is inaccurate. Meanwhile, recent studies show that LLMs excel particularly in re-ranking and challenging scenarios, rather than uniformly across all contexts.We introduce Selective LLM Guided Regularization, a model-agnostic and computation efficient framework that activates LLM based pairwise ranking supervision only when a trainable gating mechanism informing by user history length, item popularity, and model uncertainty predicts the LLM to be reliable. All LLM scoring is performed offline, transferring knowledge without increasing inference cost. Experiments across multiple datasets show that this selective strategy consistently improves overall accuracy and yields substantial gains in cold start and long tail regimes, outperforming global distillation baselines.
Paper Structure (23 sections, 3 equations, 2 figures, 3 tables, 1 algorithm)

This paper contains 23 sections, 3 equations, 2 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. LLM-Generated Soft Rankings
  5. LLM-Guided Pairwise Ranking Regularizer
  6. User-consistent pair construction.
  7. User-consistent pair construction.
  8. Selective Gating Mechanism
  9. Signals.
  10. Learnable gate.
  11. Uncertainty instantiations.
  12. Training Objective and Optimization Procedure
  13. Experimental Setup
  14. Backbone Models
  15. Datasets
  16. ...and 8 more sections

Figures (2)

  • Figure 1: Illustration of our selective LLM-guided regularization framework. Left: In the offline phase, the LLM is prompted to produce soft relevance scores Right: In the training phase, a base recommender produces prediction scores, and the LLM signals are incorporated through a pairwise ranking regularizer whose contribution is controlled by a gating function.
  • Figure 2: AUC improvements in cold-start and long-tail regimes across backbones. S-LLMR delivers the strongest boosts for cold-start users and long-tail items—often exceeding generalrelative improvement.