TSRBench
A benchmark suite for evaluating time series reasoning ability, containing multiple task types such as forecasting, anomaly detection, and numerical question answering. It is designed to measure how well models reason over temporal data rather than merely fitting it.
Timeline (last 90 days)
On the TSRBench benchmark, the model using only line graphs achieved an accuracy of 44.5%, while the dual-view model improved it to 53.2%, an improvement of nearly 9 percentage points.
The strongest specialized TSLM baseline in the paper's evaluation achieved an accuracy of 36.7% on TSRBench.
All Facts (2)
On the TSRBench benchmark, the model using only line graphs achieved an accuracy of 44.5%, while the dual-view model improved it to 53.2%, an improvement of nearly 9 percentage points.
50%UnverifiedThe strongest specialized TSLM baseline in the paper's evaluation achieved an accuracy of 36.7% on TSRBench.
50%