RobustExplain: Evaluating Robustness of LLM-Based Explanation Agents for Recommendation

TL;DR

RobustExplain addresses the robustness of LLM-based explanation agents in recommender systems under realistic noisy user histories. It introduces a perturbation taxonomy with five perturbations and a multi-dimensional robustness metric that combines semantic, keyword, structural, and length stability to quantify explanation consistency. Across four LLMs (7B–70B), the study finds moderate robustness around 0.50, with larger models achieving up to ~8% higher stability and robustness degrading only gradually with perturbation severity. The work establishes first robustness benchmarks for explanation agents and offers actionable insights for developing more trustworthy, noise-tolerant explanations at web scale.

Abstract

Large Language Models (LLMs) are increasingly used to generate natural-language explanations in recommender systems, acting as explanation agents that reason over user behavior histories. While prior work has focused on explanation fluency and relevance under fixed inputs, the robustness of LLM-generated explanations to realistic user behavior noise remains largely unexplored. In real-world web platforms, interaction histories are inherently noisy due to accidental clicks, temporal inconsistencies, missing values, and evolving preferences, raising concerns about explanation stability and user trust. We present RobustExplain, the first systematic evaluation framework for measuring the robustness of LLM-generated recommendation explanations. RobustExplain introduces five realistic user behavior perturbations evaluated across multiple severity levels and a multi-dimensional robustness metric capturing semantic, keyword, structural, and length consistency. Our goal is to establish a principled, task-level evaluation framework and initial robustness baselines, rather than to provide a comprehensive leaderboard across all available LLMs. Experiments on four representative LLMs (7B--70B) show that current models exhibit only moderate robustness, with larger models achieving up to 8% higher stability. Our results establish the first robustness benchmarks for explanation agents and highlight robustness as a critical dimension for trustworthy, agent-driven recommender systems at web scale.

Figures (4)

• Figure 1: RobustExplain evaluation framework architecture. The original user history is systematically perturbed using five perturbation types (noise injection, temporal shuffle, behavior dilution, category drift, missing values) at varying severity levels. Both original and perturbed histories are processed through LLM explanation generators, producing explanation pairs for multi-dimensional robustness comparison across semantic, keyword, structural, and length dimensions.
• Figure 2: Robustness scores across perturbation types. Drift perturbation (0.503) shows highest robustness, while shuffle presents the greatest challenge (0.499).
• Figure 3: Robustness degradation across severity levels. The gradual decline from Level 1 to Level 5 demonstrates moderate sensitivity to perturbation intensity.
• Figure 4: Robustness comparison across model sizes. LLaMA 3.1-70B achieves highest robustness (0.532), followed by Qwen2.5-14B (0.519), while 7-8B models score around 0.49.