(논문 요약) Guided Self-Evolving LLMs with Minimal Human Supervision (Paper)

핵심 내용

• R-FEW: guided Self-Play Challenger–Solver framework
  • A challenger samples human-labeled examples to guide synthetic question generation.
  • AsSolver trains on human and synthetic examples under an online, difficulty-based curriculum.
  • $\mathcal{H}_t$: historical data (previous questions, responses, document corpus)

• Related work