The path to scientific breakthroughs often follows a dual trajectory: initial, undirected exploration yielding serendipitous findings, succeeded by rigorous analysis to contextualize these discoveries within established theoretical frameworks. This paper introduces the ResearchEVO framework, an end-to-end system designed to computationally instantiate this 'discover-then-explain' paradigm.
Algorithmic Evolution Beyond Human Intuition
The Evolution Phase of ResearchEVO employs a sophisticated LLM-guided bi-dimensional co-evolutionary approach. This process simultaneously optimizes both the algorithmic logic and the overall architecture of code implementations, driven purely by fitness metrics. Crucially, this search operates without requiring any inherent understanding of the solutions it generates, enabling the exploration of novel algorithmic mechanisms. This blind search capability is central to discovering unexpected yet effective solutions.
Automated Scientific Documentation with Verifiable Grounding
Following the discovery phase, the Writing Phase of the ResearchEVO framework autonomously generates complete, publication-ready research papers. This is achieved through sentence-level retrieval-augmented generation, incorporating explicit anti-hallucination verification and automated experiment design. The system ensures that generated content is not only coherent but also factually grounded in existing literature, with a zero-fabrication policy for citations. This integrated approach bridges the gap between computational discovery and formal scientific communication.
Cross-Disciplinary Validation on Real-World Problems
The efficacy of the ResearchEVO framework was demonstrated across two challenging, cross-disciplinary scientific problems. The system was applied to Quantum Error Correction, utilizing real Google quantum hardware data, and to Physics-Informed Neural Networks. In both instances, the Evolution Phase successfully identified human-interpretable algorithmic mechanisms that were novel and not previously documented in the respective domain literature. The subsequent Writing Phase autonomously produced compilable LaTeX manuscripts, accurately situating these blind discoveries within existing theoretical contexts via RAG.