The big picture

Most benchmarks test retrieval or single-hop answers. StateBench reveals a deeper flaw: models fail to maintain an updated picture of the world.