ELT-Bench-Verified: Benchmark Quality Issues Underestimate AI Agent Capabilities

2026-03-31

AI Summary

This study addresses the significant underestimation of AI agents’ true capabilities in end-to-end ELT (Extract–Load–Transform) data pipeline construction due to inherent flaws in the existing ELT-Bench benchmark. The authors propose an Auditor-Corrector methodology that integrates large language model–driven root cause analysis with high-agreement human validation (Fleiss’ κ = 0.85) to systematically audit and rectify benchmark errors. Their analysis reveals, for the first time, that most reported failures stem from benchmark quality issues rather than agent limitations. Leveraging these insights, they construct ELT-Bench-Verified—a rigorously validated, high-quality revision of the original benchmark—and refine its evaluation logic. Re-evaluation demonstrates substantially improved AI agent performance, underscoring that reliable assessment requires concurrent attention to both model advancement and benchmark integrity, thereby establishing a more trustworthy foundation for evaluating AI-driven data engineering automation.

Abstract

Constructing Extract-Load-Transform (ELT) pipelines is a labor-intensive data engineering task and a high-impact target for AI automation. On ELT-Bench, the first benchmark for end-to-end ELT pipeline construction, AI agents initially showed low success rates, suggesting they lacked practical utility. We revisit these results and identify two factors causing a substantial underestimation of agent capabilities. First, re-evaluating ELT-Bench with upgraded large language models reveals that the extraction and loading stage is largely solved, while transformation performance improves significantly. Second, we develop an Auditor-Corrector methodology that combines scalable LLM-driven root-cause analysis with rigorous human validation (inter-annotator agreement Fleiss' kappa = 0.85) to audit benchmark quality. Applying this to ELT-Bench uncovers that most failed transformation tasks contain benchmark-attributable errors -- including rigid evaluation scripts, ambiguous specifications, and incorrect ground truth -- that penalize correct agent outputs. Based on these findings, we construct ELT-Bench-Verified, a revised benchmark with refined evaluation logic and corrected ground truth. Re-evaluating on this version yields significant improvement attributable entirely to benchmark correction. Our results show that both rapid model improvement and benchmark quality issues contributed to underestimating agent capabilities. More broadly, our findings echo observations of pervasive annotation errors in text-to-SQL benchmarks, suggesting quality issues are systemic in data engineering evaluation. Systematic quality auditing should be standard practice for complex agentic tasks. We release ELT-Bench-Verified to provide a more reliable foundation for progress in AI-driven data engineering automation.

Problem

• ELT pipeline
• benchmark quality
• AI agent evaluation
• ground truth errors
• data engineering automation

Innovation

• Auditor-Corrector methodology
• benchmark quality auditing
• ELT-Bench-Verified
• LLM-driven root-cause analysis
• agentic task evaluation