CausalForge: Automated Causal Inference Research with Lean Proof Assistant

CausalForge, a novel framework, automates theoretical research in causal inference by utilizing the Lean proof assistant to produce machine-verified proofs, thus addressing the inconsistencies associated with LLM-based reviewers. This system, created by researchers, incorporates Causalean—a foundational Lean library featuring 7,035 machine-verified declarations developed with the aid of language models under human oversight—and CausalSmith, an autonomous pipeline that identifies research topics, suggests outcomes, formalizes statements, constructs proofs, and provides artifacts for human evaluation. This initiative tackles the issue highlighted in the 2025 Bad Scientist study, which found that LLM reviewers may mistakenly approve fraudulent papers at nearly random rates. The framework is elaborated in arXiv paper 2607.22511.