Pooled accuracy summarizes performance by counting correct outcomes across a combined set of benchmark tasks and dividing by the total number evaluated. If tests have different sizes, a larger test contributes more observations to the pooled score.
That is different from taking the simple average of each benchmark's percentage. A system could do very well on a large, easy test and poorly on a smaller, difficult one yet still earn a high pooled accuracy. The choice of what to pool can therefore alter apparent rankings.
Pooling also assumes the outcomes can reasonably be scored together. Some benchmarks measure exact answers; others use human judgments, task completion or partial credit. Combining them without a consistent pass/fail rule may hide meaningful differences in the evaluation methods.
When reporting pooled accuracy, disclose the benchmark list, task counts, scoring rules and any weighting. Show per-benchmark results as well, and avoid implying that one combined number captures reliability across every kind of real-world task.