Hasty Briefsbeta

Bilingual

Can AI agents conduct open-ended AI research?

9 hours ago
  • Shadow evaluations are proposed as a new method to measure AI agents' ability to conduct open-ended AI research.
  • Agents managed to complete engineering tasks but failed to make substantial progress on research questions.
  • Five recurring failure modes were identified: poor judgment, uncreative responses, ineffective backtracking, poor resource awareness, and instruction drift.
  • A robustness check with a different model and scaffold reproduced the failures.
  • The results suggest current AI agents can handle engineering but struggle with critical parts of the research lifecycle.