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

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

Additional Submission Instructions: For the camera-ready version, please include the author names and affiliations, funding disclosures, and acknowledgements.

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

Keywords: Darwin–Gödel Discovery Machine, Risk-Aware Self-improving AI, Molecular discovery, Reinforcement learning, AI for Science

TL;DR: We introduce DGDM, a dual-loop, risk-aware AI system that improves molecular docking performance through constrained evolutionary optimization while enabling cautious pipeline self-adaptation.

Abstract:

We present the Darwin–Gödel Discovery Machine (DGDM), a dual-loop system for risk-aware self-improving AI4Science. The inner Darwinian loop evolves candidate solutions—demonstrated here with molecular ligands—via reinforcement learning–guided variation, fitness evaluation, and constraint-based retention, ensuring chemical validity and incremental improvement. Surrounding this process, an outer Gödelian loop adapts elements of the discovery pipeline itself, using confidence-based acceptance to regulate potentially harmful modifications.
In a proof-of-concept molecular docking study on four seed ligands, DGDM improves median binding affinity from $-4.457$ to $-5.422$ kcal/mol while maintaining 100\% chemical validity. These results illustrate how bounded-risk inner-loop evolution can yield scientifically meaningful gains, while highlighting the role of risk-aware acceptance in stabilizing self-directed discovery pipelines. Although preliminary in scope, this work demonstrates the feasibility of dual-loop architectures for AI-driven scientific discovery and motivates future extensions toward more robust and trustworthy self-improving AI4Science systems. A reproducibility package will be released upon publication.

Submission Number: 422