AcademiClaw: When Students Set Challenges for AI Agents

Links: arXiv · Code

AcademiClaw examines the academic-level capabilities of AI agents through tasks contributed by university students, including homework, research projects, competitions, and personal projects that current agents struggle to solve effectively.

Highlights

• Curates 80 complex bilingual tasks from 230 student-submitted candidates.
• Runs tasks in isolated Docker sandboxes with multidimensional scoring rubrics and a safety audit.
• Diagnoses capability boundaries across domains, including tasks that require GPU execution and full-stack debugging.