A new benchmark called TimeSage-EV checks whether AI agents can keep up with data that changes every month, not just data frozen in a training set.
The team built TimeSage-EV around 60 real institutional scenarios across six domains, using 1,485 scenario-period question-and-answer pairs collected from February 2023 through May 2026. Data drops follow the same schedules the source institutions actually use: monthly, weekly, daily, or irregular. At each release, an AI agent receives the time series and any accompanying reports, then has to identify the current state, summarize the data, and reason about what comes next. The next real release, withheld until then, serves as the answer key. The researchers also tested TimeSage-1.0, a self-evolving agent built for the benchmark that keeps a reusable library of analytical skills.
Static benchmarks let a model look sharp by leaning on patterns baked into old training data, which says little about handling a report that lands next week. TimeSage-EV's rolling design forces agents to use only the evidence available at each moment, and the frontier models tested still stumbled, with recurring failures in respecting that cutoff, pulling in outside context, and adapting as new numbers arrived.
The benchmark refreshes monthly with a public leaderboard, so any claim of beating it will have to keep beating it, not just pass once.