Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Benchmark Leakage Trap: Can We Trust LLM-based Recommendation?

Mingqiao Zhang, Qiyao Peng, Yumeng Wang, Chunyuan Liu, Hongtao Liu

TL;DR

This paper identifies and investigates a previously overlooked issue: benchmark data leakage in LLM-based recommendation, which reveals that data leakage acts as a critical, previously unaccounted-for factor in LLM-based recommendation, which could impact the true model performance.

Abstract

The expanding integration of Large Language Models (LLMs) into recommender systems poses critical challenges to evaluation reliability. This paper identifies and investigates a previously overlooked issue: benchmark data leakage in LLM-based recommendation. This phenomenon occurs when LLMs are exposed to and potentially memorize benchmark datasets during pre-training or fine-tuning, leading to artificially inflated performance metrics that fail to reflect true model performance. To validate this phenomenon, we simulate diverse data leakage scenarios by conducting continued pre-training of foundation models on strategically blended corpora, which include user-item interactions from both in-domain and out-of-domain sources. Our experiments reveal a dual-effect of data leakage: when the leaked data is domain-relevant, it induces substantial but spurious performance gains, misleadingly exaggerating the model's capability. In contrast, domain-irrelevant leakage typically degrades recommendation accuracy, highlighting the complex and contingent nature of this contamination. Our findings reveal that data leakage acts as a critical, previously unaccounted-for factor in LLM-based recommendation, which could impact the true model performance. We release our code at https://github.com/yusba1/LLMRec-Data-Leakage.

Benchmark Leakage Trap: Can We Trust LLM-based Recommendation?

TL;DR

This paper identifies and investigates a previously overlooked issue: benchmark data leakage in LLM-based recommendation, which reveals that data leakage acts as a critical, previously unaccounted-for factor in LLM-based recommendation, which could impact the true model performance.

Abstract

The expanding integration of Large Language Models (LLMs) into recommender systems poses critical challenges to evaluation reliability. This paper identifies and investigates a previously overlooked issue: benchmark data leakage in LLM-based recommendation. This phenomenon occurs when LLMs are exposed to and potentially memorize benchmark datasets during pre-training or fine-tuning, leading to artificially inflated performance metrics that fail to reflect true model performance. To validate this phenomenon, we simulate diverse data leakage scenarios by conducting continued pre-training of foundation models on strategically blended corpora, which include user-item interactions from both in-domain and out-of-domain sources. Our experiments reveal a dual-effect of data leakage: when the leaked data is domain-relevant, it induces substantial but spurious performance gains, misleadingly exaggerating the model's capability. In contrast, domain-irrelevant leakage typically degrades recommendation accuracy, highlighting the complex and contingent nature of this contamination. Our findings reveal that data leakage acts as a critical, previously unaccounted-for factor in LLM-based recommendation, which could impact the true model performance. We release our code at https://github.com/yusba1/LLMRec-Data-Leakage.
Paper Structure (22 sections, 9 equations, 3 figures, 6 tables)

This paper contains 22 sections, 9 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Large Language Models
  4. LLM-based Recommendation
  5. Methodology
  6. Overview
  7. Leakage Data Construction
  8. Injecting leaked data
  9. Loss definition
  10. Contaminated adaptation (Dirty LLM)
  11. LoRA parameterization and what is updated
  12. Justification for LoRA as a controlled proxy
  13. Experiment
  14. Experiment settings
  15. Datasets
  16. ...and 7 more sections

Figures (3)

  • Figure 1: Overview of the Experimental Framework. The workflow illustrates the construction of a Mixed Leakage Corpus through strategic sampling of In-Domain (ID) and Out-Of-Domain (OOD) data. The Clean Path represents the baseline evaluation using a frozen Base LLM, while the Dirty Path simulates benchmark contamination via LoRA injection. By comparing the Clean and Dirty Recommenders, we identify the Triple Effect of Leakage: spurious performance gains, stability, or degradation.
  • Figure 2: AUC under different data injection methods
  • Figure 3: UAUC under different data injection methods