Evolutionary Intelligence for Scientific Discovery: From Evolutionary Computation to Cumulative Discovery Systems

Chao Wang, Lingling Li, Fang Liu, Licheng JiaoarXiv 2026 · arXiv

TL;DR

This paper proposes a paradigm shift from traditional Evolutionary Computation (EC) to "Evolutionary Intelligence" (EI) for autonomous scientific discovery. By systematically retaining and transferring experience across evolutionary cycles—rather than just searching for a single optimal solution—EI transforms isolated trial-and-error trajectories into cumulative scientific insight.

What Problem Does This Paper Address?

Artificial intelligence is rapidly moving from passive data analysis to driving autonomous, closed-loop scientific discovery in fields like molecular design, protein engineering, and automated laboratories. Historically, Evolutionary Computation (EC) has provided a strong foundation for navigating these complex, gradient-free search spaces using population-based search.

However, traditional EC treats discovery as a task-specific, black-box optimization problem. It focuses heavily on refining candidates for a predefined problem and discards the rest of the search history once an optimal candidate is found. The authors argue that this approach misses the true nature of science, which is inherently cumulative. Valuable data from failed experiments, intermediate candidate lineages, and complex multi-dimensional feedback are typically lost, preventing AI systems from building long-term scientific knowledge or transferring insights to new problems.

Methods

As a review and perspective paper, the authors do not introduce a new empirical algorithm, but instead formalize a conceptual framework to define and guide the transition from EC to EI. They introduce two core analytical tools:

  1. A Five-Dimensional Analytical Framework: This framework categorizes and evaluates scientific AI systems by asking five interconnected questions:
  2. What evolves? (Scientific targets, computational components, or discovery processes)
  3. How do candidates change? (Traditional mutations vs. foundation-model or human-guided variations)
  4. Why are candidates selected? (Multi-dimensional trade-offs, physical feasibility, and domain value)
  5. Where does feedback originate? (Computational proxies, physical constraints, or expert judgment)
  6. When does evolution occur? (Design/training stages vs. inference/experimental stages)
  7. The Evolutionary Intelligence Cycle: A conceptual pipeline illustrating how an AI system transforms raw search trajectories into structured knowledge. The cycle consists of five phases: Experience Retention (archiving failures and logs) $\rightarrow$ Knowledge Representation (building surrogate models or scientific priors) $\rightarrow$ Knowledge Utilization (adjusting search strategies and proposal distributions) $\rightarrow$ Knowledge Transfer (reusing skills and models across tasks) $\rightarrow$ Scientific Insight (extracting interpretable design rules and hypotheses).
The transition from Evolutionary Computation (EC) to Evolutionary Intelligence (EI) for scientific discovery.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Failures are Features, Not Bugs: In the EI paradigm, failed trials and sub-optimal candidates are systematically archived as critical evidence. This history maps out physically impossible or operationally constrained regions of the search space, preventing the AI from repeating past mistakes.
  • Broadening the Evolving Object: Evolution is no longer restricted to the final scientific target (e.g., a molecular structure). EI systems can concurrently evolve the underlying machine learning models, prompt templates, hypotheses, and even the experimental execution workflows themselves.
  • Foundation Models as Evolutionary Operators: Large language models (LLMs) fundamentally change how candidates vary. Instead of relying on random mutations, LLMs leverage encoded scientific priors to propose physically or logically plausible variations in code, hypotheses, and experimental plans.
The five-dimensional framework used to analyze and design Evolutionary Intelligence systems.

Key Results

While the paper does not present new quantitative experimental data, it provides a comprehensive taxonomy unifying recent high-profile breakthroughs under the EI umbrella—such as DeepMind’s FunSearch (for algorithmic program discovery) and various closed-loop mobile robotic chemists.

The authors map out how EI expands every core operation of traditional EC. For instance, instead of reducing evaluation to a simple scalar fitness value, EI expands feedback to incorporate multi-fidelity computational proxies, real-world physical constraints, and human expert oversight.

Limitations and Future Work

The authors identify several critical bottlenecks that the AI4Science community must address to fully realize EI:

  • Longitudinal Evaluation: Current evaluation frameworks only measure single-cycle candidate refinement. New benchmarks are required to track how effectively an AI system accumulates and transfers knowledge over successive, long-term discovery cycles.
  • Process Traceability: When foundation models drive candidate variation, tracing the exact evolutionary lineage of a discovery becomes difficult. Robust logging and human-in-the-loop verification are essential to turn raw model outputs into trustworthy scientific evidence.
  • Simulation-to-Experiment Gaps: Bridging the gap between fast, cheap, but imperfect computational simulations and slow, expensive physical laboratory experiments remains a major operational challenge.
  • Shared Infrastructure: There is a distinct lack of standardized databases for experience reuse. The community needs shared data formats to systematically record and share failed experiments, modification histories, and trained models across different domains.

Summary updated

Abstract

Artificial intelligence (AI) is shifting scientific discovery from task-specific workflows towards autonomous systems that organize exploration with experimental and human feedback in open-ended candidate spaces. Evolutionary computation (EC) provides a computational basis for feedback-driven discovery because population-based search can maintain diverse scientific candidates while steering exploration through accumulated evidence. However, EC predominantly focuses on candidate refinement for predefined problems, whereas cumulative discovery requires experience retention. To bridge this gap, this review introduces evolutionary intelligence (EI) for scientific discovery. EI characterizes scientific AI systems that sustain exploration by linking candidate refinement with experience retention across evolutionary cycles. We introduce a five-dimensional analytical framework that asks what evolves, how candidates change, why candidates are selected, where feedback originates, and when evolution occurs. This framework clarifies how EI transforms isolated search trajectories into cumulative scientific insight. We further demonstrate this paradigm across diverse discovery modes, from evolving concrete scientific entities to orchestrating automated research workflows. Finally, we identify critical bottlenecks regarding evaluation, process traceability, and shared infrastructure, providing a concrete roadmap for advancing the transition from EC to EI in scientific discovery.
BibTeX
@misc{wang2026evolutionary,
  author = {Chao Wang and Lingling Li and Fang Liu and Licheng Jiao},
  title = {{Evolutionary Intelligence for Scientific Discovery: From Evolutionary Computation to Cumulative Discovery Systems}},
  year = {2026},
  eprint = {2607.09025},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.09025},
}
Figure