SCOPE

"SCOPE: Self-Play via Co-Evolving Policies for Open-Ended Tasks" — data-free self-play in which a Challenger generates document-grounded tasks, a Solver answers them via multi-turn retrieval, and a frozen self-judge grades the exchange; sits at the open-endedness + RL intersection. Clean Edinburgh work (Kwan, Gema, Leang, Minervini), with code and models released via EdinburghNLP.