Test-Time Self-Play: Scaling without a Verifier

Researchers introduce Test-Time Self-Play (TTSP), a test-time scaling method that improves large language model performance on hard target problems without relying on a verifier. Evaluations on a difficulty-calibrated subset of LiveCodeBench v6 show that TTSP outperforms existing methods across various verifier strengths, achieving a 35 percent absolute improvement in mean at eight in the no-verifier limit. Open article →