Self-Harmony: Learning to Harmonize Self-Supervision and Self-Play in Test-Time Reinforcement Learning

Researchers introduce Self-Harmony, a new test-time reinforcement learning framework that improves AI model accuracy by having models solve problems and rephrase questions simultaneously. The method uses harmonic mean aggregation instead of majority voting to select stable answers, achieving state-of-the-art results across 28 of 30 reasoning benchmarks without requiring human supervision.

• →Self-Harmony framework uses a single model in dual roles as both problem solver and question reframer to improve reliability.
• →The method replaces majority voting with harmonic mean aggregation to avoid spurious but popular answers.
• →Achieved first place results in 28 of 30 test settings across multiple reasoning benchmarks.
• →The approach requires no human supervision or auxiliary models, making it highly practical.
• →Demonstrated zero training failures across all experiments, showing exceptional stability and robustness.