Pinned
Today we’re releasing LEAPBench, a benchmark for how efficiently models learn from feedback over many rounds.
Most evals only score whether a model eventually reaches a good solution. LEAPBench scores how many tries it took to get there.
Two systems can reach the same optimum



