Title: TSQAgent: Enhancing Time Series Data Quality Assessment through Specialized Agentic Reasoning

Abstract:

Evaluating the quality of time series (TS) data is a critical yet complex task, primarily due to the multidimensional nature of quality metrics. While large language models (LLMs) have recently shown potential in this domain through pairwise comparisons and per-dimension evaluations, current methods are constrained by two main factors: they depend on manually defined quality dimensions and rely exclusively on text-based reasoning. Consequently, it remains unclear whether LLMs can autonomously detect truly pertinent quality dimensions or execute grounded, quantitative comparisons.

To address this gap, we introduce TSQBench, a specialized benchmark designed to test LLMs on two advancing competencies: (i) the ability to comprehend and pinpoint relevant quality dimensions, and (ii) the capacity to conduct quality comparisons within those specific dimensions. Our findings indicate that existing LLMs face significant difficulties in both identifying appropriate dimensions and performing evidence-based quality comparisons.

In response to these challenges, we present TSQAgent, an innovative agentic reasoning framework for rating TS quality. This framework operates through three distinct, collaborative roles: the Perceiver, which selects specific dimensions for focus; the Inspector, which conducts quantitative analysis on a dimension-by-dimension basis; and the Adjudicator, which synthesizes and refines the final assessment. TSQAgent incorporates a novel reasoning strategy that empowers the system to identify and prioritize the most significant quality dimensions. Furthermore, we have developed an agent workflow integrated with external analytical tools to facilitate precise, quantitative comparisons across selected dimensions.

Experimental results, conducted on both our new benchmark and eleven real-world datasets, demonstrate that TSQAgent significantly boosts LLM capabilities in understanding quality and performing quantitative comparisons. These advancements effectively improve quality-aware data selection, resulting in superior downstream performance and greater data efficiency.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC