(블로그 요약) Universe: Scale Real-World Verifiable Environments to Millions (paper)

핵심 내용

• Builder agent
  • code base 에서 bug fix.
  • evaluation.sh 생성.
• Hacking Detector
  • evaluation.sh 가 제대로 된 것인지 검토.
• Verifier agent
  • buggy branch 에서는 테스트 실패, fixed branch 에서는 테스트 통과 하는지 체크.
• 데이터셋
  • train: ~1M high-quality PRs (filtered from 33.3 M pull requests in 2021–2025 of GitHub)
    • high quality: PRs: linked with issue(s) without excessive file changes
  • benchmark: 320 pull requests randomly sampled from GitHub
• 모델
  • Qwen-Next-80B-A3B
  • Qwen3-Max-Thinking