Learning to Evaluate Before Improving: Automatic Rubric Induction for Automatic Research Agents
AutoSciRub has research agents write an executable grading rubric before they start the research task.
The framework breaks vague scientific instructions into atomic goals, grounds them in literature and visible task data, then turns them into verifiable criteria. Those criteria guide the agent’s analyses and experiments, and later flag unmet requirements for targeted revision. In the reported tests, it improved results across ResearchClawBench and a 20-task subset of AstaBench E2E Discovery while maintaining or increasing task completion. ArXiv · AI/CL/LG's note
The framework breaks vague scientific instructions into atomic goals, grounds them in literature and visible task data, then turns them into verifiable criteria. Those criteria guide the agent’s analyses and experiments, and later flag unmet requirements for targeted revision. In the reported tests, it improved results across ResearchClawBench and a 20-task subset of AstaBench E2E Discovery while maintaining or increasing task completion. ArXiv · AI/CL/LG's note
score 5