Yufeng Wang introduced the FirstResearch framework on July 6, 2026, aiming to enhance the auditability of research questions proposed by large language models (LLMs) in scientific discovery. This framework addresses the challenge of verifying the initial research questions generated by LLMs, which often lack transparency in their underlying assumptions.
Understanding the FirstResearch Framework
The FirstResearch framework is designed to provide clear and structured research question formation for scientific LLM agents. Its core component, the Research Question Certificate, documents essential elements such as primitive definitions, assumptions, and hypotheses. This structured approach ensures that researchers can inspect the proposed questions before proceeding with further analysis.
Key features of the Research Question Certificate include:
- Primitive definitions
- Assumptions and mechanisms
- Tensions or contradictions
- Falsifiable hypotheses
- Minimal decisive tests
- Failure update rules
Performance Evaluation of FirstResearch
In a comparative study involving ten research topics, FirstResearch demonstrated superior performance against controlled prompt-level baselines inspired by existing AI models such as AI co-scientist and AI Scientist-v2. Under a primary evaluation protocol known as DeepSeek-blind-judge, FirstResearch achieved a score of 4.86/5, significantly higher than the strongest baseline's score of 4.38/5.
