SCOPE: Self-Play via Co-Evolving Policies for Open-Ended Tasks

Researchers introduced SCOPE, a self-play framework for training open-ended language models without curated prompts or external supervision.

It pairs a Challenger that creates document-grounded tasks with a Solver that answers using retrieval, then grades results with a frozen self-judge.

On three 7–8B instruction-tuned models, SCOPE improved open-ended benchmarks by up to 10.4 points and held-out short-form QA by up to 13.8 points.

It outperformed a supervised GRPO_data baseline, suggesting self-play can co-evolve task creation and answering policies effectively.