Researchers have released a benchmark designed to test AI models on time series data that looks like what engineers actually encounter, not the clean, evenly spaced kind.
IRTS-ToolBench contains 1,700 questions across 10 task types and 13 domains, all built around irregular time series: observations that arrive asynchronously, missing values that carry meaning rather than noise, and sensors sampling at different rates. Most existing Time Series Question Answering benchmarks assume data arrives at regular intervals, which is closer to a textbook than a factory floor. The new benchmark is designed to be used as a standalone tool, with standardized inputs and a reproducible evaluation protocol, and the code is publicly available. The researchers frame the gap as fundamental: we largely don't know how well large language models and AI agents hold up when the data is messy.
That matters because irregularity is the default in high-stakes deployments. Medical monitors drop packets. Industrial sensors go offline. Financial feeds stutter. If a benchmark can't measure how an agent handles those conditions, the accuracy numbers it produces are optimistic by design. A shared standard for this specific failure mode closes a real gap in how the field measures itself.
It is, of course, still a benchmark. AI benchmarks have a habit of getting saturated the moment they gain traction. The more useful question is whether IRTS-ToolBench ends up in standard evaluation suites or quietly becomes a citation in related-work sections.