New AI Training Paradigm Turns Open-Ended Tasks Into Games, Boosting Performance Across Writing, Summarization, and Math

A groundbreaking AI training method called RLSVR transforms open-ended tasks like writing and summarization into competitive multi-agent games, achieving win rates above 75% while simultaneously boosting mathematical reasoning by up to 9% — all without requiring human judges or labeled data.

• A new training paradigm called RLSVR (Reinforcement Learning with Self-Verifiable Rewards) transforms open-ended tasks into multi-agent game environments that automatically generate verifiable reward signals, extending RLVR beyond math and code to domains like summarization and creative writing.
• The framework is instantiated as SpyRL, a multi-agent self-play system inspired by 'Who Is the Spy?' where agents receive asymmetric information, complete a shared task, and vote to identify a designated spy, with voting outcomes serving as fully verifiable rewards without requiring any external judge or ground-truth labels.
• SpyRL achieves strong results on Qwen3-8B with 75.4% and 77.3% A/B win rates on summarization and creative writing respectively, while also improving mathematical reasoning on Qwen3-4B and Qwen3-8B by 8.97% and 6.16% across seven benchmarks, significantly outperforming competing methods like R-Zero and Absolute Zero.