Autonomous scientific research agents are increasingly used for end-to-end workflows, from literature review to report generation. However, open-ended tasks often lack clear specifications for analyses, methods, and success criteria, leading agents to miss important steps or draw unsupported conclusions.
To address this, researchers introduce AutoSciRub, an evaluation-first framework that induces a task-specific executable rubric before research execution. The rubric decomposes underspecified instructions into atomic scientific goals, grounds them in relevant literature and data, and synthesizes specific, actionable, and verifiable criteria.
During execution, the rubric guides experiments and analyses; during revision, rubric-guided verification identifies unmet criteria and enables targeted refinement of the report and supporting artifacts. This makes implicit experimental and evidential requirements explicit.
On ResearchClawBench, AutoSciRub improved all tested configurations, with average gains of 2.08 points across three backbone LLMs under a fixed Codex harness and 2.95 points across three agent harnesses with a fixed DeepSeek-V4-Flash backbone. On a 20-task subset of AstaBench E2E Discovery, it achieved an average improvement of 16.8 points across three agent harnesses while maintaining or increasing task completion.
The results suggest that evaluation-first guidance is an effective and generalizable control mechanism for autonomous scientific research. Code is available on GitHub.