On-Policy Self-Training: Data-Free Self-Improvement in Large Language Models

Hrih News: A new paper, On-Policy Self-Training (OPST), proposes improving language models without human-curated questions or answers. Seeded only by coarse topic descriptors, the model generates its own question-answer pairs, filters them through pseudo-verification and a consensus gate, and trains via on-policy self-distillation from a privileged EMA teacher.