Skip to main content
Back

Sun Yat-Sen's TSQAgent Uses AI Agents to Cut Noisy Training Data by 50%

TSQAgent uses a three-role LLM framework with external tools to automatically identify and measure time series data quality, improving downstream forecasting and classification by selecting better training data

Paper, Full Access
  • TSQAgent is a three-role LLM framework (Perceiver, Inspector, Adjudicator) that rates time series data quality using adaptive dimension selection and external analytical tools.
  • The paper introduces TSQBench, a new benchmark revealing that current LLMs consistently fail at both identifying relevant quality dimensions and making quantitative comparisons.
  • The Perceiver role is trained with GRPO (reinforcement learning from quality comparison rewards), improving dimension identification precision by 20-25 percentage points over baselines.
  • External tools let the Inspector compute actual statistics (autocorrelation, spectral density, SNR) rather than relying on verbal reasoning, grounding comparisons in real numbers.
  • Experiments on 11 real-world datasets show that selecting the top 50% of training data by TSQAgent quality scores consistently reduces forecasting error and improves classification accuracy.
  • The related open-source TSRating codebase is available; TSQAgent code availability was not confirmed at time of publication.

Garbage in, garbage out. That principle is especially brutal in time series modeling, where a single noisy sensor reading or a corrupted segment can quietly degrade a forecasting model for months. Knowing which data is high-quality, and why, has historically required either expensive human review or brittle heuristics. A new paper from researchers at Sun Yat-Sen University and the National University of Singapore proposes TSQAgent, an agentic LLM framework that can autonomously rate time series data quality with quantitative, tool-grounded reasoning.

The problem with letting LLMs judge data quality

Time series data is everywhere: electricity grids, hospital monitors, financial exchanges, traffic sensors. The quality of time series data is critical to model performance, as low-quality data can negate the benefits of technical advancements, even potentially leading to detrimental performance. In practice, time series datasets are frequently marred by pervasive quality issues, including missing values, sensor-failure-induced corruption, and irregular sampling.

Recent work has tried to use LLMs as quality judges by prompting them to compare pairs of time series or score them along predefined dimensions like trend, frequency, amplitude, and pattern. The predecessor to this work, TSRating, showed that LLMs can indeed distinguish high- from low-quality time series with over 89% accuracy on synthetic benchmarks. But that approach had a fundamental blind spot: the quality dimensions were manually chosen by researchers, and the comparisons were purely text-based, with no access to actual numerical analysis tools.

TSQAgent's authors asked a harder question: can LLMs figure out which quality dimensions actually matter for a given dataset, and can they make quantitative comparisons rather than just verbal ones? The answer, before their intervention, was a clear no.

Building a benchmark to expose the gap

Before proposing a fix, the team built TSQBench, a dedicated benchmark designed to probe two specific capabilities in LLMs: