FirstResearch Enhances Auditable Scientific Question Generation by LLMs.

Large Language Model (LLM) systems are increasingly assisting in scientific discovery, from ideation to experiment planning. However, the initial research questions proposed by these agents often lack transparency, making it difficult for scientists to audit their underlying mechanisms, assumptions, or falsifiability. This new framework, called FirstResearch, aims to address this by making LLM-generated scientific questions more inspectable. FirstResearch introduces a core artifact: a structured Research Question Certificate. This certificate meticulously records primitive definitions, underlying assumptions, a proposed mechanism model, any identified tensions or contradictions, a falsifiable hypothesis, a minimal decisive test, and a rule for updating based on failure. By requiring the LLM agent to generate this certificate alongside the question, the entire derivation process becomes transparent and auditable before any downstream execution. Preliminary results, judged by other LLMs (DeepSeek and Gemini-2.5-Flash), show that FirstResearch significantly outperforms controlled prompt-level baselines inspired by other AI co-scientist systems. The certificate-centered approach was identified as the strongest component, with certificate-only scoring achieving very high marks. While these results are preliminary and use LLM judges, they strongly suggest that explicit derivation constraints are a promising method for enhancing the audibility of LLM-generated scientific questions.