A benchmark for irregular time‑series question answering has been released.
The IRTS-ToolBench suite contains 1,700 questions across 10 task types and 13 domains. It is built for LLM‑based agents and offers a standard input format and evaluation script. The dataset targets the kinds of gaps—async observations, informative missing values, and mixed sampling rates—that real deployments present but that existing TSQA tests ignore.
This matters because most prior TSQA work assumes neatly sampled data, giving a false sense of capability. By forcing models to confront realistic irregularities, the benchmark can expose weaknesses in current tool‑grounded reasoning pipelines and guide more robust architectures.
If the community adopts IRTS‑ToolBench, we may see a shift from toy‑like time‑series tasks toward evaluations that reflect field conditions, much as image‑recognition benchmarks did a decade ago.