“Current large language models do not perform well on Diplomacy benchmarks.”