Scientific Algorithm Discovery by Augmenting AlphaEvolve with Deep Research
Gang Liu, Yihan Zhu, Jie Chen, Meng Jiang
2025-10-08
Summary
This paper introduces DeepEvolve, a new system designed to help scientists by automatically creating and improving computer algorithms for solving problems in fields like chemistry and biology.
What's the problem?
Current attempts to use large language models (LLMs) for scientific discovery have weaknesses. Some systems just try to evolve algorithms internally, but they quickly hit a limit because they don't have enough outside knowledge. Others generate ideas without checking if they actually work in the real world, leading to impractical solutions. Essentially, existing systems either lack grounding in reality or get stuck in endless, unproductive tweaking.
What's the solution?
DeepEvolve combines the strengths of both approaches. It uses LLMs to research existing knowledge, then proposes changes to an algorithm. Crucially, it doesn't just suggest changes – it actually *implements* those changes in code, tests them, and uses the results to refine its approach in a continuous loop. This process of research, implementation, and testing happens repeatedly, leading to consistently better algorithms. It can edit code across multiple files and systematically find and fix errors.
Why it matters?
DeepEvolve is important because it offers a more reliable way to use LLMs for scientific algorithm discovery. By connecting research with practical testing and improvement, it avoids the pitfalls of previous methods and provides a framework for making real progress in various scientific fields, demonstrated through improvements across nine different benchmarks.
Abstract
Large language models hold promise as scientific assistants, yet existing agents either rely solely on algorithm evolution or on deep research in isolation, both of which face critical limitations. Pure algorithm evolution, as in AlphaEvolve, depends only on the internal knowledge of LLMs and quickly plateaus in complex domains, while pure deep research proposes ideas without validation, resulting in unrealistic or unimplementable solutions. We present DeepEvolve, an agent that integrates deep research with algorithm evolution, uniting external knowledge retrieval, cross-file code editing, and systematic debugging under a feedback-driven iterative loop. Each iteration not only proposes new hypotheses but also refines, implements, and tests them, avoiding both shallow improvements and unproductive over-refinements. Across nine benchmarks in chemistry, mathematics, biology, materials, and patents, DeepEvolve consistently improves the initial algorithm, producing executable new algorithms with sustained gains. By bridging the gap between unguided evolution and research without grounding, DeepEvolve provides a reliable framework for advancing scientific algorithm discovery. Our code is available at https://github.com/liugangcode/deepevolve.