Agentic Artificial Intelligence is transforming recommender systems through agent-based architectures that integrate diverse data modalities and advanced reasoning. While recent work emphasizes agent design, the quality of input data remains largely overlooked. This paper aims to provide the basic concepts to assess how data quality impacts recommendation process and user satisfaction from diferent perspectives in agentic settings.
Every day, users face vast and ever-growing catalogs of items, struggling with information overload,
which degrades their experience and threatens platform revenues. Recommender Systems (RSs) [
Recommendation algorithms have undergone many innovations and advancements, with the latest
being Agentic AI architectures, based on Large Language Models (LLMs)-empowered agents, which have
gained traction in recommender systems research [
While the performance of traditional recommender systems is known to depend heavily on input
data quality [
To bridge this gap, we propose evaluation criteria that aim to decouple the influence of input data quality from algorithmic performance. These principles aim to support more transparent evaluation practices, foster fairer comparisons across systems, and encourage thoughtful data and prompt design in future research on agentic recommender systems.
The rise of LLM-powered agentic systems in recommendation tasks [
While recent works such as MACRec [
We focus on two key and underexplored dimensions of this issue:
2. the quality of the prompts used to define agent roles, behavior, and coordination.
In contrast to traditional recommender systems, where noisy or incomplete data directly undermines performance [18], agentic systems may compensate for poor input through the reasoning abilities and pretrained knowledge of LLMs [19]; however, this does not make data quality irrelevant, rather, it suggests the need for rethinking it as a minimal threshold that enables efective agent collaboration.
To study this, we first introduce Relative Data Quality (RDQ) for measuring performance variation across diferent versions of the same dataset (e.g., difering in annotation quality, modality richness, or encoding), while keeping architecture and prompting fixed. Comparing across diferent datasets would introduce confounding factors like domain or sparsity, making such comparisons less informative. RDQ instead isolates the impact of specific data quality dimensions under controlled settings, allowing us to decouple the contribution of data quality from that of the algorithm itself. For example, experiments could assess how changes in image resolution, text length, or preprocessing impact system performance, and to what extent LLMs’ reasoning capabilities make them robust to such variations.
Prompting strategies assume a central role in agentic systems, where the prompts define the roles
and the allowed actions of each agent [
It is worth noticing that absolute evaluation is not meaningful in this context, as removing prompts or input data would render the system non-functional. Instead, we advocate for assessing relative performance variations under controlled and fixed conditions. This approach ofers more actionable insights by isolating the impact of specific elements, such as structured metadata or image inputs, on agents’ reasoning and decision-making capabilities. These observations can inform system design, highlight critical dependencies, and support more transparent development and debugging processes.
This paper laid the groundwork for evaluating how input data quality influences agentic recommender systems, highlighting the often-overlooked role of data, particularly in multimodal, agentic settings. By analyzing the impact of diferent modalities and prompt configurations, we aim to encourage more transparent and data-aware evaluation practices. Future work may explore modality-specific quality assessment and adaptive agent routing based on evaluated data quality, positioning data quality as a central concern in agentic RSs.
[17] C. Bauer, L. Chen, N. Ferro, N. Fuhr, A. Anand, T. Breuer, G. Faggioli, O. Frieder, H. Joho, J. Karlgren, J. Kiesel, B. P. Knijnenburg, A. Lipani, L. Michiels, A. Papenmeier, M. S. Pera, M. Sanderson, S. Sanner, B. Stein, J. R. Trippas, K. Verspoor, M. C. Willemsen, Manifesto from dagstuhl perspectives workshop 24352 - conversational agents: A framework for evaluation (CAFE), CoRR abs/2506.11112 (2025). [18] Y. Li, Q. Zhao, C. Lin, J. Su, Z. Zhang, Who to align with: Feedback-oriented multi-modal alignment in recommendation systems, in: SIGIR, ACM, 2024, pp. 667–676. [19] L. Wu, Z. Zheng, Z. Qiu, H. Wang, H. Gu, T. Shen, C. Qin, C. Zhu, H. Zhu, Q. Liu, H. Xiong, E. Chen, A survey on large language models for recommendation, World Wide Web (WWW) 27 (2024) 60.