The Darwin–Gödel Discovery Machine: Toward Bounded-Risk Self-Improving AI4Science

Xuening Wu, Xinhang Zhang, Yanlan Kang, Qianya Xu, Honggang Wang, Zeping Chen

Track: Track 1: Original Research/Position/Education/Attention Track

Keywords: Darwin–Gödel Machine, Self-improving AI, Molecular discovery, Reinforcement learning, PAC bounds, AI4Science

TL;DR: This paper proposes the Darwin–Gödel Discovery Machine, a dual-loop framework that couples reinforcement-learning–guided molecular evolution with PAC-style statistical safeguards for bounded-risk self-improving AI4Science.

Abstract: We present the Darwin–Gödel Discovery Machine (DGDM), a dual-loop framework for bounded-risk self-improving AI4Science. The inner Darwinian loop evolves candidate solutions—demonstrated here with molecular ligands—through reinforcement learning (RL) guided variation, fitness evaluation, and constraint-based retention, ensuring validity and incremental improvement. The outer Gödelian loop adapts the discovery pipeline itself, governed by statistical safeguards (PAC–style) that limit harmful modifications. In a proof-of-concept docking study on four seed ligands, DGDM improved median binding affinity from -4.457 to -5.422 kcal/mol while preserving 100% chemical validity. These results illustrate how bounded-risk inner-loop evolution can yield scientifically meaningful advances, while motivating future extensions of the outer loop for trustworthy pipeline optimization. Although preliminary in scope, this work highlights the potential of dual-loop architectures to push the boundaries of AI in scientific discovery while explicitly accounting for risk. Looking ahead, RL strategies and large language models with domain-grounded retrieval offer natural mechanisms to enrich inner-loop adaptation and outer-loop self-improvement, advancing the vision of trustworthy, self-improving AI4Science. An anonymized reproducibility package will be released to facilitate community feedback.

Submission Number: 422