「By actively collecting information, the agent can update its beliefs and make better decisions under uncertainty. To evaluate a model’s ability to reason while actively acquiring information, we draw inspiration from the concept of Interactive Proofs in computational complexity theory (Goldwasser et al , 2019) and propose a unified evaluation paradigm, which we call Interactive Benchmarks.」という行動しながら答えを見出すタイプのベンチマーク。現実的に重要なタスク。（汎用モデルで）このような動作が可能になってきているのも感慨深いものがある。