Grouping search trajectories by acquired knowledge rather than surface states enables more reliable credit assignment in long-horizon search, letting smaller models (8B) match larger ones (32B+) on knowledge-intensive tasks.
KnowGPO improves how AI agents learn to search for answers by tracking what knowledge they've discovered rather than just looking at surface-level observations. Instead of giving credit only at the end, it breaks down each question into required knowledge pieces and rewards agents for discovering them step-by-step, making learning more reliable in complex search tasks.