PREreview del UniDataBench: Evaluating Data Analytics Agents Across Structured and Unstructured Data

This paper introduces UniDataBench, a comprehensive benchmark designed to evaluate data analytics agents across heterogeneous data sources, including relational databases, CSV/Excel files, and NoSQL systems. The authors correctly identify a critical limitation in existing benchmarks, which tend to focus on single data modalities and fail to reflect the multi-source, messy data environments common in real-world business analytics. By grounding the benchmark in sanitized industry analysis reports, the work increases its practical relevance while addressing privacy and sensitivity concerns.

A key contribution of the paper is the unified evaluation framework that assesses agents’ ability to explore diverse data formats, reason across sources, and generate meaningful summaries and recommendations. The benchmark moves beyond isolated query execution to emphasize end-to-end analytical capability, which is essential for realistic assessment of data analytics agents. The proposed agent, ReActInsight, further demonstrates how large language model–based agents can autonomously decompose analytical goals, discover cross-source relationships, and generate self-correcting code to extract actionable insights.

While the benchmark and agent are well motivated, broader comparisons across alternative agent architectures would further strengthen the evaluation. Overall, this work represents a significant step toward more realistic and rigorous assessment of data analytics agents in enterprise-like settings and provides a valuable foundation for future research in this area.

Competing interests

The author declares that they have no competing interests.

Use of Artificial Intelligence (AI)

The author declares that they did not use generative AI to come up with new ideas for their review.