Paper deep dive
Evolutionary Intelligence for Scientific Discovery: From Evolutionary Computation to Cumulative Discovery Systems
Chao Wang, Lingling Li, Fang Liu, Licheng Jiao
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 7/13/2026, 2:28:02 AM
Summary
This review paper introduces Evolutionary Intelligence (EI) as a paradigm shift from traditional Evolutionary Computation (EC) to cumulative scientific discovery. EI links candidate refinement with experience retention across evolutionary cycles, enabling autonomous AI systems to transform isolated search trajectories into structured scientific insights. The authors propose a five-dimensional analytical framework to evaluate how evolving objects, variation sources, selection criteria, feedback environments, and evolutionary time scales interact. The paper highlights applications in autonomous laboratories, foundation models, and agentic science, while identifying bottlenecks in evaluation, traceability, and shared infrastructure.
Entities (8)
Relation Signals (6)
Evolutionary Computation → providesbasisfor → Scientific Discovery
confidence 95% · 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.
Evolutionary Intelligence → bridgesgapbetween → Evolutionary Computation
confidence 94% · 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.
Five-dimensional Framework → analyzes → Evolutionary Intelligence
confidence 93% · 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.
Experience Retention → enables → Cumulative Discovery
confidence 92% · cumulative discovery requires experience retention. ... EI treats the entire search trajectory, including failed trials and candidate lineages, as scientific evidence to be structured and reused.
Autonomous Laboratories → implement → Evolutionary Intelligence
confidence 90% · Concurrently, autonomous laboratories and scientific agents integrate exploration with experimental and human feedback, enabling subsequent candidates to adapt dynamically to accumulated evidence.
Foundation Models → drives → Source of Variation
confidence 88% · Foundation-model-assisted evolutionary methods further expand this scope beyond final scientific targets to fundamental research components. These systems actively evolve prompts, code, and hypotheses, alongside reusable skills.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2607.09025v1
- Canonical: https://arxiv.org/abs/2607.09025v1
Trouble viewing inline? Open PDF directly →
Full Text
52,904 characters extracted from source content.
Expand or collapse full text
Evolutionary Intelligence for Scientific Discovery: From Evolutionary Computation to Cumulative Discovery Systems Chao Wang 1 , Lingling Li 1 , Fang Liu 1 , Licheng Jiao 1* 1* School of Artificial Intelligence, Xidian University, No. 2 South Taibai Road, Xi’an, 710071, Shaanxi, China. *Corresponding author(s). E-mail(s): lchjiao@mail.xidian.edu.cn; Contributing authors: xiaofengxd@126.com; llli@xidian.edu.cn; f63liu@163.com; Abstract Artificial intelligence (AI) is shifting scientific discovery from task-specific work- flows towards autonomous systems that organize exploration with experimental and human feedback in open-ended candidate spaces. Evolutionary computa- tion (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 evo- lutionary 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 analyti- cal framework that asks what evolves, how candidates change, why candidates are selected, where feedback originates, and when evolution occurs. This frame- work clarifies how EI transforms isolated search trajectories into cumulative scientific insight. We further demonstrate this paradigm across diverse discov- ery 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. Keywords: Evolutionary computation, AI for Science, scientific discovery, autonomous experimentation, foundation models, self-evolving agents 1 arXiv:2607.09025v1 [cs.NE] 10 Jul 2026 1 Introduction Artificial intelligence (AI) increasingly serves as a general-purpose tool for scientific discovery, transitioning from task-specific automation to systems that organize scien- tific exploration [1, 2]. By leveraging experimental feedback and human interaction, AI systems can explore scientific problems through discovery cycles. This shift is evi- dent across disciplines, where AI actively investigates candidate spaces in molecular and materials design [3, 4], protein engineering [5], and automated laboratories [6, 7], moving beyond passive data analysis. Feedback-driven discovery explores open-ended candidate spaces where boundaries are often undefined and emerge as evidence accumulates [1, 8]. Because experimental feedback is costly and physically constrained, every evaluated candidate, including failures, constitutes valuable scientific evidence [9, 10]. Consequently, the evaluation history can reveal hidden structures within the candidate space and support diverse exploratory behaviours [8, 11]. Evolutionary computation (EC) provides a foundational computational frame- work to support these discovery cycles [12, 13]. Rather than relying on a single-point search, EC maintains populations of scientific candidates that are iteratively updated by evolutionary operators according to feedback. This population-based approach is particularly well suited to open-ended candidate spaces characterized by weak or absent gradients, noisy feedback, and dynamically evolving boundaries [8, 14–16]. By steering exploration toward regions supported by accumulated evidence, EC sustains the creation of diverse scientific candidates across discovery cycles. This approach encompasses various common algorithms, such as classical genetic algorithm, evolution strategies [17–20], natural evolution strategies [21, 22], and quality-diversity search methods [23–26]. Conventionally, EC has advanced AI for scientific discovery primarily by refin- ing candidates for well-defined scientific problems. Extensive research has focused on designing specialized evolutionary operators to steer this candidate refinement [17– 19, 21]. However, cumulative discovery relies on more than just finding the best candidate. The entire search trajectory, including failed trials and complex feedback, provides experience that guides later searches [24, 26–28]. Therefore, evolutionary search serves not just to solve single problems, but to accumulate experience across iterative discovery cycles [29, 30]. To capture this transition, 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. Within this framework, populations of scientific candidates are updated by evolu- tionary operators and evaluated via experimental or human feedback. Specifically, these evaluated candidates and their corresponding feedback are retained in struc- tured forms to guide subsequent exploration. Unlike single-task search methods that primarily maximize sample efficiency for isolated problems [15, 31, 32], EI treats the entire search trajectory, including failed trials and candidate lineages, as scientific evi- dence to be structured and reused. By bridging candidate refinement with experience retention, EI provides a structured framework for cumulative discovery [27, 29, 30]. 2 Recent AI systems increasingly embed evolutionary cycles into scientific discovery. Algorithm-discovery platforms, such as FunSearch and AlphaEvolve, refine candidate programs through automated evaluation mechanisms [33, 34]. Foundation-model- assisted evolutionary methods further expand this scope beyond final scientific targets to fundamental research components. These systems actively evolve prompts, code, and hypotheses, alongside reusable skills [8, 35–39]. Concurrently, autonomous lab- oratories and scientific agents integrate exploration with experimental and human feedback, enabling subsequent candidates to adapt dynamically to accumulated evi- dence [7, 9, 40–44]. The emergence of such autonomous research systems is increasingly recognized as agentic science [45–49]. As evolutionary search begins to reshape these research components, evaluation metrics need to expand beyond mere can- didate quality to explicitly incorporate scientific reliability and human oversight [11, 39, 50–52]. This review examines the transition from EC to EI in scientific discovery. We first introduce a five-dimensional analytical framework designed to analyze diverse scientific AI systems, from foundation-model-assisted program search to closed-loop autonomous laboratories. This framework clarifies how the EI cycle transforms iso- lated search experiences into structured scientific insights. We further demonstrate how this paradigm operates across diverse discovery modes, from evolving concrete scientific entities to orchestrating automated research workflows. Finally, we outline key challenges regarding evaluation, process traceability, and shared infrastructure, providing a concrete roadmap to advance this transition. Ultimately, this perspective reframes EC from a task-level search tool into an organizing principle for cumulative scientific discovery. 2 From evolutionary computation to evolutionary intelligence EC provides a foundational framework for population-based search in scientific dis- covery. Rather than relying on single-point updates, EC maintains populations of candidates through evolutionary operators that organize reproduction, evaluate outcomes against feedback, and drive selection. Adaptive strategies and archival mechanisms retain historical experience to influence subsequent search trajectories [13, 17, 18, 20, 21]. This organizational logic is well suited to open-ended scientific candidate spaces. In such environments, search often proceeds before boundaries, objectives, or feedback processes are fully defined, aligning with broader biological perspectives on adaptation and the evolution of physical artifacts [12, 53]. By preserv- ing diverse candidate lineages and leveraging quality-diversity principles, EC sustains robust exploration even under weak gradients or noisy feedback [14, 15, 24–26, 54]. While EC provides the search mechanism, many traditional applications treat sci- entific discovery as black-box optimization over fixed representations. Unlike these approaches, EI broadens the scope of the evolving object. Candidate populations in EC need not be stored exclusively as explicit sets. While genetic algorithms [18] and genetic programming [19] traditionally maintain explicit candidate representations, natural evolution strategies [21, 22] and estimation-of-distribution algorithms [55–57] 3 Evolutionary Computation (EC) Reproduction MutationCrossover + Selection Elites . . . Candidate Refinement Against Feedback Population . . . Evaluation Against Feedback Environment Feed back Intera ction Evolution Evolutionary Intelligence (EI) for Scientific Discovery Cumulative Scientific Discovery Scientific Discovery Systems Scientific Targrts . . . Research Components . . . Discovery Processes Scientific Candidate Computational Experimental Human Feedback Evaluation Against Feedback Candidate Refinement Cycle To Fig. 1 From evolutionary computation (EC) to evolutionary intelligence (EI) for scientific discovery. EC organizes population-based search through evolutionary operators. EI embeds this organization in scientific discovery systems where evolution can act on scientific targets, research components, or discovery processes. The transition from EC to EI shifts the emphasis from candidate refinement against feedback to cumulative scientific discovery. update probabilistic sampling distributions derived from evaluated candidates. In sci- entific discovery, evolution can act on scientific targets [4, 5], foundational research components [33, 35, 36, 58, 59], up to the discovery processes themselves [9, 40, 41, 60]. For instance, rather than merely refining a molecular structure, an EI system might concurrently evolve the machine learning model predicting its properties, alongside the experimental batching policy used to test it [43, 44, 61]. This shift from fixed rep- resentations to dynamic, multi-level targets allows scientific AI systems to organize exploration flexibly, whether through explicit candidate pools and learned genera- tive models, alongside probabilistic proposal distributions. However, as the evolving objects become more complex and scientifically grounded, the evaluation mechanism needs to evolve beyond simple scalar fitness assignment. Scientific feedback expands the role of evaluation beyond simple ranking to incor- porate diverse physical and computational constraints. In EC practices, evaluation is typically reduced to a scalar objective value used to rank candidates [17, 18, 62]. In scientific discovery, however, feedback encompasses various types of signals. These range from data-driven computational proxies, such as predictive models and uncer- tainty estimates [3–5, 15], to physical constraints derived from laboratory experiments [7, 9, 10, 40], alongside human-driven reproducibility checks and expert judgement [1, 2, 11]. Consequently, these diverse signals shape not only which candidates are selected, but also how the underlying models, constraints, and experimental priorities are adjusted for the subsequent discovery cycle. These signals generate rich scientific evidence across every cycle, which raises the key question of how such evidence is preserved and utilized over time. 4 Table 1 From evolutionary computation to evolutionary intelligence for scientific discovery. DimensionEvolutionary computationEvolutionary intelligence for scientific discovery Primary rolePopulation-based searchCumulative scientific discovery PopulationCandidate sets or sampling distri- butions Scientific targets, research compo- nents, or discovery processes ReproductionMutation,crossover,distribu- tional sampling, and program variation Evolutionary variation, learned generation, foundation-model gen- eration, and human-guided modi- fication Feedback-based evaluation Fitness functions, objective val- ues, or simulations Computational,experimental, and human feedback Selectionand update Candidate survival and distribu- tion update Updated candidates, proposal dis- tributions, models, constraints, and experimental priorities Archiveand record Elites, non-dominated solutions, diverse behaviours, and search states Successful candidates, failures, lineages, model states, experimen- tal records, and annotations Experience reten- tion and transfer Adaptation within a search loop and transfer across related tasks Experience reuse across evolution- ary cycles and transfer across tasks, models, skills, or experi- ments Ultimately, the integration of expanded evolving objects and diverse scientific feed- back requires a shift in the temporal scale of experience retention, marking a key transition from isolated search to cumulative scientific discovery. Traditional adaptive or transfer methods typically confine experience utilization within a single search loop [63] or between closely related tasks [64, 65]. By contrast, EI systematically archives the search trajectory, treating failed trials, candidate lineages, and experimental logs as valuable scientific evidence. This structured record reshapes candidate generation and selection pressures in subsequent cycles, enabling knowledge transfer across dis- tinct tasks and scientific domains [28, 66–68], as well as simulation-to-experiment boundaries [65]. Recent foundation-model-assisted agents [8, 33–36] and autonomous laboratories [9, 40, 41] exemplify this paradigm. In these systems, validation, trace- ability, and human oversight operate as core elements of the EI framework rather than external checks [11, 52, 69]. Fig. 1 and Table 1 summarize this transition from EC to EI for scientific discovery. Fig. 1 illustrates how population-based search becomes embedded within feedback-driven discovery systems, while Table 1 compares their core operations. 5 What Evolves？ Scientific Targrts . . . Research Components . . . Discovery Processes A Five-dimensional Framework for Evolutionary Intelligence for Scientific Discovery Evolving Object How Candidates Change? Generative Models Foundation Models . . . Human Modification Source of Variation . . . Why Candidates are Selected? Trade-offs Selection Criterion Physical Feasibility Domain-specific Value Where Feedback Originates？ Computational Proxies Feedback Environment Physical Constraints Expert Judgement When Evolution Occurs？ Design and Training . . . Time Scale of Evolution Inference and Experiments Time Scales . . . . . . . . . . . . . . . . . . Fig. 2 A five-dimensional framework for evolutionary intelligence. This framework analyzes scientific AI systems by linking five analytical questions to their corresponding system components: the evolving object (what evolves), the source of variation (how candidates change), the selection criterion (why candidates are selected), the feedback environment (where feedback originates), and the time scale of evolution (when evolution occurs). 3 A five-dimensional framework for evolutionary intelligence EI for scientific discovery can be analyzed through five interconnected questions: what evolves, how candidates change, why candidates are selected, where feedback origi- nates, and when evolution occurs. These questions correspond to five components of a discovery system: the evolving object, the source of variation, the selection criterion, the feedback environment, and the time scale of evolution [17]. These components interact closely. The feedback environment influences the selection criteria, which in turn shape the mechanisms of variation, while the time scale determines how accu- mulated experience is retained and transferred across cycles [28]. This framework, illustrated in Fig. 2, provides a structured approach to analyze how isolated search trajectories are transformed into cumulative scientific insight. What evolves. Many EC applications typically target fixed representations, rang- ing from numerical vectors to static programs and predefined policies [17–19, 70, 71]. For scientific discovery, EI broadens this scope across multiple levels. The evolving object may encompass concrete scientific targets, research components, or discovery processes. Specifically, targets range from molecular structures and protein sequences to complex experimental recipes [4, 5, 72, 73]. Research components include compu- tational models and data representations [74–76], as well as prompt templates and testable hypotheses [35, 36, 77]. Furthermore, discovery processes themselves can evolve, including experimental planning protocols, tool-use routines, and human-AI collaborative workflows [7, 40, 41, 60]. This broader view is valuable because these objects frequently overlap in practice. A scientific AI system often adapts multiple components concurrently within a single cycle. For instance, an autonomous materials platform might simultaneously refine candidate compounds, update underlying surro- gate models, and adjust experimental batching policies [15, 43, 44, 61]. Similarly, the evolution of reusable skills, such as automating data curation and literature retrieval, 6 functions as a dynamic research component that supports the broader discovery workflow [39, 60]. How candidates change. EC algorithms primarily drive candidate variation through operators such as mutation and crossover, as well as distributional sam- pling and program transformation [18, 19, 21, 55]. In EI for scientific discovery, candidate variation is also driven by learned generative models, foundation models, and human-guided modification. For instance, large foundation models can propose or revise complex programs and testable hypotheses, and refine prompt templates and experimental plans [8, 33–36, 78, 79]. Domain-specific scientific constraints and prior knowledge can help guide this variation, restricting it to physically or logically plausible regions of the candidate space [50, 76]. Why candidates are selected. Many EC implementations rely on scalar fit- ness values or predefined objective functions to rank and propagate candidates [17, 18, 62]. In scientific discovery, however, selection criteria expand to encompass multi-dimensional trade-offs, physical feasibility, and domain-specific scientific value. Rather than prioritizing a single metric, selection pressures balance competing objec- tives, such as predictive accuracy, computational cost, and experimental safety [80, 81]. Furthermore, mechanisms that preserve novelty and diversity penalize premature con- vergence, keeping alternative hypotheses and unexplored regions of the candidate space viable [23–27]. These criteria adapt as feedback accumulates, while human judgement and expert oversight can steer selection toward scientifically promising regions that lack immediate computational validation [32, 82, 83]. Consequently, selection in EI functions not only to filter candidates, but also to steer the search trajectory toward evolving scientific priorities. Where feedback originates. EC practices typically derive evaluation feed- back from predefined, scalar fitness functions or controlled simulated environments [17, 18, 62]. In scientific discovery, however, the feedback environment includes var- ious types of signals. These range from computational proxies, such as predictive models and uncertainty estimates [3–5, 84], to physical constraints derived from lab- oratory experiments [7, 9, 10, 40, 85], and human-driven reproducibility checks with expert judgement [1, 2, 11]. These signals do more than merely rank candidates; they influence how models, search constraints, and experimental priorities are adjusted for subsequent discovery cycles. Consequently, the feedback environment in EI serves as a direct link between the computational search and physical experiments. When evolution occurs. The temporal scale at which evolutionary operations are applied influences whether a system remains confined to local search or sup- ports systematic cumulative discovery. In scientific AI systems, evolution operates at multiple time scales, depending on the latency and cost of feedback. At the design and training stages, evolutionary mechanisms shape model architectures [58, 86] and search configurations [59, 87], and adjust proposal distributions during the learning process [21, 88]. At the inference and experimental stages, foundation-model systems can search over reasoning paths [77, 79], algorithmic programs [33, 34], and testable scientific hypotheses [7, 36, 37], while laboratory feedback refines candidate batches and experimental priorities [9, 40, 89]. Across time scales, archived priors and adaptive search strategies enable knowledge transfer across distinct tasks and scientific domains 7 [28, 66–68]. Consequently, the timing of evolution is more than a scheduling detail; it determines how retained experience shapes future discovery cycles. Collectively, these five dimensions provide a framework for comparing diverse sci- entific AI systems across different domains and technical implementations [1, 47]. By examining how a system combines its evolving objects, variation mechanisms, and feedback environments, we can determine whether it is limited to isolated search or capable of systematic cumulative discovery. This perspective shifts the focus from indi- vidual algorithms to how the system learns over time. Consequently, it provides a basis for analyzing the EI cycle, which integrates these dimensions to transform isolated search trajectories into scientific insight. 4 The evolutionary intelligence cycle: from experience retention to scientific insight The role of search trajectories. EI expands the value of search trajectories. Many EC applications primarily focus on identifying a set of high-performing candidates or a Pareto front [24, 25, 27, 80]. For cumulative discovery, however, the entire search trajectory can serve as a valuable foundation for generating scientific insight. Rather than merely recording final outcomes, EI systems frequently capture detailed evidence. Failed trials often reveal physical and operational constraints, while candidate lin- eages document specific modifications that preserve or degrade function. Furthermore, experimental records can highlight systematic discrepancies between computational surrogates and physical reality [9, 15, 16, 40]. By archiving these complete search histories, EI translates isolated search trajectories into a structured experience that informs subsequent discovery cycles. This transformation is governed by the EI cycle (Fig. 3), which systematically retains, represents, utilizes, and transfers experience to generate scientific insight. Experience retention. This mechanism differs from common EC practices of archiving primarily elite solutions or non-dominated fronts [17, 18, 23, 24]. In scien- tific discovery, EI systems often employ broader archival protocols that record more aspects of the search process. Rather than discarding suboptimal outcomes, these pro- tocols frequently incorporate failed trials, intermediate evolutionary lineages, and raw experimental logs, and human annotations with contextual metadata [6, 9, 10, 40]. By structuring these diverse records into a unified historical record, EI helps preserve a broader range of scientific evidence. This extensive retention provides valuable raw material for subsequent knowledge representation, converting isolated data points into a coherent and traceable scientific record. Knowledge representation. Raw archival records often require transformation into usable and interpretable forms to guide future exploration. In this phase, EI systems convert accumulated search history into structured representations. These representations range from data-driven probabilistic surrogate models that map can- didate features to evaluation indicators [15, 90, 91], to topological mappings of the search space that reveal hidden basins of attraction [24, 27], and domain-specific sci- entific priors derived from physical laws or expert knowledge [1–3]. This step connects low-level data to higher-level reasoning. By recording both successes and failures, the 8 Evolutionary Intelligence Cycle: From Experience Retention to Scientific Insight 2. Knowledge Representation Probabilistic Models Search Space Mappings Scientific Priors 1. Experience Retention Failed Trials Candidate Lineages Experimental Logs 5. Scientific Insight Structure--Property Relationships Design Rules Testable Mechanistic Hypotheses 3. Knowledge Utilization Search Strategies Selection Pressures Programs and Hypotheses 4. Knowledge Transfer Tasks and Domains Skills Simulation-to- Experiment Knowledge Base Systematic Validation and Traceability Fig. 3 From experience retention to scientific insight. Evolutionary intelligence (EI) transforms disconnected search trajectories into a coherent knowledge base that informs subsequent discovery cycles. Experience retention captures and structures diverse scientific records. Knowledge rep- resentation converts these raw records into interpretable representations. Knowledge utilization actively employs this retained memory to adjust search strategies, selection pressures, programs, and hypotheses, reshaping how candidates are generated, evaluated, and selected. Knowledge trans- fer enables this archived experience to accumulate scientific value across diverse scientific contexts. Ultimately, this cumulative process aims to produce scientific insight, grounded in systematic val- idation and traceability. system builds a structured history that guides future searches toward viable options and prevents repeated mistakes. Knowledge utilization. This mechanism uses retained experience to steer the ongoing search process. Rather than relying on static rules, EI systems adjust their search strategies using the structured representations built in the previous phase. For instance, data-driven surrogates and uncertainty estimates can guide the genera- tion of new candidates, directing evolutionary variation toward regions of the search space that are promising or uncertain [15, 91, 92]. Topological mappings and novelty metrics change selection pressures, penalizing premature convergence and encourag- ing the exploration of diverse behaviors [23–25]. Furthermore, foundation models use encoded scientific priors to propose programs and valid hypotheses, reducing reliance on random mutation [7, 33, 35, 36]. Through these mechanisms, knowledge utilization shifts the search from trial-and-error to evidence-driven inquiry, balancing open-ended exploration and targeted refinement as new feedback arrives. Knowledge transfer. This mechanism enables cumulative discovery by reusing experience across tasks, rather than limiting search to single runs or closely related problems [28, 63, 64]. In EI systems, structured records and adaptive strategies are reused across new scientific contexts. Rather than starting from scratch, these systems use archived priors to accelerate exploration in distinct tasks and scientific domains [1, 7, 41, 47]. This transfer also bridges computational and physical boundaries. For instance, heuristics and data-driven models learned in simulations can be transferred 9 to guide costly physical experiments, helping to connect simulations with physical experiments [15, 65, 68]. Similarly, experimental protocols and design principles can propagate across different length scales or material classes [3, 4, 76]. Systems also extract and reuse modular skills. When they find effective algorithms or prompt tem- plates, they archive these structures for future tasks [35, 36, 39, 60]. Systems can then combine these saved skills to solve new problems. By reusing past results, each discovery cycle builds directly on previous knowledge rather than starting from scratch. Scientific insight. This phase shifts the focus from finding high-performing can- didates to generating interpretable scientific knowledge. While many search processes conclude with a single best candidate or a Pareto front, cumulative discovery sys- tems extract generalized principles from the entire search trajectory. These insights include explicit structure–property relationships, design rules, or testable mechanistic hypotheses that explain why certain candidates succeed while others fail [1, 2, 11]. Because these insights are based on archived failures, diverse search paths, and phys- ical validation, they are highly reliable and traceable [9, 16]. Over time, this new scientific insight feeds back into the broader research community, improving human understanding and providing initial priors for future autonomous discovery systems. 5 Modes of EI-enabled scientific discovery To demonstrate the generality of the EI framework, we analyze scientific AI systems based on their evolving object rather than their specific application domains. This approach shows a progression in the scope of evolution: from concrete scientific enti- ties and computational models, through symbolic reasoning processes, to automated research workflows. This classification shows that while specific candidate spaces and feedback mechanisms vary across disciplines, the EI architecture remains consistent [1, 47]. Evolving concrete scientific entities. EI operates on physical or chemical tar- gets, including molecules [5], proteins [72, 73], and materials [4]. Because experimental feedback in these domains is often costly and sparse, systems frequently rely on data- driven surrogate models to guide exploration [15]. EI systems in this context often archive failed synthesis routes and negative characterization data. Rather than discard- ing these outcomes, the system uses them to identify physically feasible regions. This retained experience allows subsequent search cycles to avoid inaccessible areas and focus on viable spaces, as demonstrated in autonomous materials platforms [43, 44, 61]. Modifying computational models and representations. EI operates on the computational tools used to model scientific entities. Rather than searching directly over physical targets, these systems evolve neural architectures [58, 86] and molecu- lar representations [75, 76, 93–95], and data-driven surrogate models [15]. Adapting these representations can reshape the search space, making previously inaccessible regions easier to explore. At this level, the system records how different representa- tions capture physical or chemical constraints. Because this experience can transfer across related tasks, it enables the system to select representations that favor physi- cally plausible solutions. For example, integrating foundation models into protein and 10 materials design uses pretrained representations to guide the search toward viable biological or chemical structures [50, 96–99]. Evolving symbolic reasoning processes. At a higher level, EI systems target symbolic structures, including executable code, mathematical expressions [33, 34], and mechanistic hypotheses. Unlike physical experiments, the feedback for these symbolic entities can be automatically verified, often relying on massive parallel computing. This automated verification allows systems to evaluate a large volume of candidates. Instead of discarding code that fails a test suite, EI systems often archive these errors. By analyzing these failures, the system can identify recurring logical mistakes or extract partially correct subroutines to guide subsequent modifications [8, 79]. Over successive cycles, these archived code components and corrections enable the system to construct more complex algorithms or mathematical proofs, often resulting in human-readable solutions [36, 37]. Orchestrating automated research workflows. Beyond modifying single com- ponents, EI systems can target entire research workflows. These workflows integrate multiple scientific operations, from planning to experimental execution [7, 41]. At this level, the evolving objects are the execution plans for these tasks. Feedback typically evaluates whether the complete workflow successfully produces an outcome, such as a verified hypothesis. Because this requires completing all steps, the feedback is often sparser and more delayed than component-level signals [42, 44]. To manage these pro- cesses, EI systems archive successful multi-step protocols. By retaining these execution plans, the system can reuse them to automate complex scientific tasks [39, 51, 60, 100]. In practice, a single scientific workflow often combines multiple modes. For instance, an automated system might evolve the code used to analyze data while simultaneously synthesizing physical materials. Regardless of the objects being modi- fied, the core mechanism remains the same: the system archives past results to guide subsequent steps. As EI systems integrate these modes into complex workflows, eval- uation must look beyond the final output. Rather than measuring only the quality of a single result, assessments need to evaluate the reliability of the entire process and the system’s capacity to accumulate experience over time. 6 Evaluation, challenges and roadmap Evaluating EI for scientific discovery requires criteria that go beyond the perfor- mance of isolated candidates. While candidate quality is a necessary baseline, EI is distinguished by its ability to retain experience, transfer knowledge, and produce interpretable insights across evolutionary cycles. Achieving this cumulative discovery presents several key challenges: defining evaluation metrics, improving process trace- ability, bridging simulation-to-experiment gaps, and developing shared databases for experience reuse. Evaluating cumulative discovery beyond candidate refinement. Current evaluation frameworks primarily measure the quality of final candidates within isolated discovery cycles. Such single-cycle metrics are often insufficient to distinguish whether a system accumulates transferable knowledge or merely overfits to a specific task [1, 11]. Consequently, this limitation makes it difficult to separate short-term candidate 11 refinement from systematic cumulative discovery. Addressing this gap requires a shift toward longitudinal metrics. Rather than relying on single-run performance, these metrics need to track how knowledge is utilized and transferred across successive cycles, as well as the emergence of interpretable scientific insights [7, 101]. Traceability of evolutionary processes. When foundation models generate variations, tracing how a specific candidate was derived is often difficult. Without a clear record of the evolutionary process, scientists cannot independently verify or reproduce the results [52, 69]. To address this, systems must record the complete history of candidate modifications, including both successful steps and failed attempts. Additionally, involving human experts to review these records helps ensure that the generated candidates align with known physical or logical rules, turning raw model outputs into trustworthy scientific evidence [11, 79]. Feedback reliability and simulation-to-experiment gaps. The EI cycle depends heavily on feedback, which is often costly, sparse, and noisy. Real-world exper- iments are slow and expensive, while computational simulations are fast but frequently simplify physical laws [40, 89]. Relying solely on either source can mislead the search. To address this, systems must combine simulations with experiments. They need to measure the gap between simulated predictions and real experimental outcomes, using this difference to continuously correct their models and avoid physically impossible regions [82, 83]. Shared infrastructure for experience reuse. Many current EI systems operate independently, generating large amounts of data that are difficult for other researchers to reuse. The lack of shared data platforms hinders the transfer of knowledge across different scientific domains, length scales, and simulation-to-experiment boundaries [3, 4]. To address this, the community needs standardized databases that systemati- cally record failed experiments, modification histories, and trained models. Developing common data formats and sharing standards would allow new EI systems to begin their searches using previously accumulated experience, enabling cumulative scientific discovery across the broader research community [1, 47]. 7 Outlook The transition to EI for scientific discovery shifts the focus from automating single tasks to accumulating knowledge over time. Future scientific AI systems will likely integrate foundation models with automated experiments through evolutionary cycles [1, 47]. In these systems, every evaluated candidate and failed experiment contributes to a shared record. This accumulated history allows subsequent searches to build on existing scientific knowledge rather than starting from scratch. As these systems become more autonomous, the role of human researchers will shift toward setting overall goals and validating results. Scientists remain essential for applying domain expertise and interpreting unexpected experimental outcomes [11, 52]. Therefore, the success of EI should not be judged solely by computational speed, the number of candidates generated, and accuracy. Instead, it depends on the system’s ability to produce experimentally validated insights that withstand scientific scrutiny and help researchers understand the underlying mechanisms. 12 References [1] Wang, H. et al. Scientific discovery in the age of artificial intelligence. Nature 620, 47–60 (2023). [2] Hey, A. J., Tansley, S., Tolle, K. M. et al. The fourth paradigm: data-intensive scientific discovery Vol. 1 (Microsoft research Redmond, WA, 2009). [3] Butler, K. T., Davies, D. W., Cartwright, H., Isayev, O. & Walsh, A. Machine learning for molecular and materials science. Nature 559, 547–555 (2018). [4] Merchant, A. et al. Scaling deep learning for materials discovery. Nature 624, 80–85 (2023). [5] Jumper, J. et al. Highly accurate protein structure prediction with alphafold. nature 596, 583–589 (2021). [6] King, R. D. et al. The automation of science. Science 324, 85–89 (2009). [7] Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624, 570–578 (2023). [8] Lehman, J. et al.Evolution through large models.arXiv preprint arXiv:2206.08896 (2022). [9] MacLeod, B. P. et al. Self-driving laboratory for accelerated discovery of thin- film materials. Science Advances 6, eaaz8867 (2020). [10] Granda, J. M., Donina, L., Dragone, V., Long, D.-L. & Cronin, L. Controlling an organic synthesis robot with machine learning to search for new reactivity. Nature 559, 377–381 (2018). [11] Messeri, L. & Crockett, M. J. Artificial intelligence and illusions of understanding in scientific research. Nature 627, 49–58 (2024). [12] Eiben, A. E. & Smith, J. From evolutionary computation to the evolution of things. Nature 521, 476–482 (2015). [13] Fogel, D. B. An introduction to simulated evolutionary optimization. IEEE transactions on neural networks 5, 3–14 (1994). [14] Salimans, T., Ho, J., Chen, X., Sidor, S. & Sutskever, I. Evolution strategies as a scalable alternative to reinforcement learning. arXiv preprint arXiv:1703.03864 (2017). [15] Jin, Y., Wang, H., Chugh, T., Guo, D. & Miettinen, K. Data-driven evolutionary optimization: An overview and case studies. IEEE Transactions on Evolutionary Computation 23, 442–458 (2018). 13 [16] Kudela, J. A critical problem in benchmarking and analysis of evolutionary computation methods. Nature Machine Intelligence 4, 1238–1245 (2022). [17] Holland, J. H. Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence (MIT press, 1992). [18] Booker, L. B., Goldberg, D. E. & Holland, J. H. Classifier systems and genetic algorithms. Artificial intelligence 40, 235–282 (1989). [19] Koza, J. R. Genetic programming as a means for programming computers by natural selection. Statistics and computing 4, 87–112 (1994). [20] Beyer, H.-G. & Schwefel, H.-P. Evolution strategies–a comprehensive introduc- tion. Natural computing 1, 3–52 (2002). [21] Hansen, N. & Ostermeier, A. Completely derandomized self-adaptation in evolution strategies. Evolutionary computation 9, 159–195 (2001). [22] Wierstra, D. et al. Natural evolution strategies. The Journal of Machine Learning Research 15, 949–980 (2014). [23] Lehman, J. & Stanley, K. O. Abandoning objectives: Evolution through the search for novelty alone. Evolutionary computation 19, 189–223 (2011). [24] Mouret, J.-B. & Clune, J. Illuminating search spaces by mapping elites. arXiv preprint arXiv:1504.04909 (2015). [25] Pugh, J. K., Soros, L. B. & Stanley, K. O. Quality diversity: A new frontier for evolutionary computation. Frontiers in Robotics and AI 3, 40 (2016). [26] Lehman, J., Meyerson, E., El-Gaaly, T., Stanley, K. O. & Ziyaee, T. Evo- lution and the knightian blindspot of machine learning.arXiv preprint arXiv:2501.13075 (2025). [27] Cully, A., Clune, J., Tarapore, D. & Mouret, J.-B. Robots that can adapt like animals. Nature 521, 503–507 (2015). [28] Tan, K. C., Feng, L. & Jiang, M. Evolutionary transfer optimization-a new frontier in evolutionary computation research. IEEE Computational Intelligence Magazine 16, 22–33 (2021). [29] Wang, C. et al. When large language models meet evolutionary algorithms: Potential enhancements and challenges. Research 8, 0646 (2025). [30] Gao, H.-a. et al. A survey of self-evolving agents: What, when, how, and where to evolve on the path to artificial super intelligence. arXiv preprint arXiv:2507.21046 (2025). 14 [31] Frazier, P. I.A tutorial on bayesian optimization.arXiv preprint arXiv:1807.02811 (2018). [32] Settles, B.Active learning literature survey.Tech. Rep., University of Wisconsin–Madison (2009). [33] Romera-Paredes, B. et al. Mathematical discoveries from program search with large language models. Nature 625, 468–475 (2024). [34] Novikov, A. et al. Alphaevolve: A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131 (2025). [35] Fernando, C., Banarse, D., Michalewski, H., Osindero, S. & Rockt ̈aschel, T. Promptbreeder: Self-referential self-improvement via prompt evolution. arXiv preprint arXiv:2309.16797 (2023). [36] Agrawal, L. A. et al.Gepa: Reflective prompt evolution can outperform reinforcement learning. arXiv preprint arXiv:2507.19457 (2025). [37] Oh, J. et al. Discovering state-of-the-art reinforcement learning algorithms. Nature 648, 312–319 (2025). [38] Zhang, B., Wang, C. & Wu, K.Evom: Agentic meta-evolution of actor- critic architectures for reinforcement learning. arXiv preprint arXiv:2606.26327 (2026). [39] Xia, P. et al. Skillrl: Evolving agents via recursive skill-augmented reinforcement learning. arXiv preprint arXiv:2602.08234 (2026). [40] Burger, B. et al. A mobile robotic chemist. Nature 583, 237–241 (2020). [41] Xi, Z. et al. Agentgym: Evolving large language model-based agents across diverse environments. arXiv preprint arXiv:2406.04151 (2024). [42] Zhang, J., Hu, S., Lu, C., Lange, R. & Clune, J. Darwin godel machine: Open-ended evolution of self-improving agents. arXiv preprint arXiv:2505.22954 (2025). [43] Ghareeb, A. E. et al. A multi-agent system for automating scientific discovery. Nature 1–3 (2026). [44] Gottweis, J. et al. Accelerating scientific discovery with co-scientist. Nature 1–3 (2026). [45] Wei, J. et al. From ai for science to agentic science: A survey on autonomous scientific discovery. arXiv preprint arXiv:2508.14111 (2025). 15 [46] Ren, S. et al. Towards scientific intelligence: A survey of llm-based scientific agents. arXiv preprint arXiv:2503.24047 (2025). [47] Chen, Q. et al. Ai4research: A survey of artificial intelligence for scientific research. arXiv preprint arXiv:2507.01903 (2025). [48] Chen, Z. et al. Scienceagentbench: Toward rigorous assessment of language agents for data-driven scientific discovery. arXiv preprint arXiv:2410.05080 (2024). [49] Assump ̧c ̃ao, H., Ferreira, D., Campos, L. & Murai, F. Codeevolve: An open source evolutionary coding agent for algorithm discovery and optimization. arXiv preprint arXiv:2510.14150 (2025). [50] Akiba, T., Shing, M., Tang, Y., Sun, Q. & Ha, D. Evolutionary optimization of model merging recipes. Nature Machine Intelligence 7, 195–204 (2025). [51] Xia, P. et al. Agent0: Unleashing self-evolving agents from zero data via tool- integrated reasoning. arXiv preprint arXiv:2511.16043 (2025). [52] Lu, C. et al. Towards end-to-end automation of ai research. Nature 651, 914–919 (2026). [53] Miikkulainen, R. & Forrest, S.A biological perspective on evolutionary computation. Nature Machine Intelligence 3, 9–15 (2021). [54] Sarkar, B. et al.Evolution strategies at the hyperscale.arXiv preprint arXiv:2511.16652 (2025). [55] Larra ̃naga, P. & Bielza, C. Estimation of distribution algorithms in machine learning: A survey. IEEE Transactions on Evolutionary Computation 28, 1301– 1321 (2024). [56] Rubinstein, R. The cross-entropy method for combinatorial and continuous optimization. Methodology and computing in applied probability 1, 127–190 (1999). [57] Mockus, J. The application of bayesian methods for seeking the extremum. Towards global optimization 2, 117 (1998). [58] Stanley, K. O., Clune, J., Lehman, J. & Miikkulainen, R. Designing neural networks through neuroevolution. Nature Machine Intelligence 1, 24–35 (2019). [59] Jaderberg, M. et al. Population based training of neural networks. arXiv preprint arXiv:1711.09846 (2017). [60] Wang, G. et al. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023). 16 [61] Chu, Y. et al. Programmable rna translation through deep learning-driven ires discovery and de novo generation. Nature Machine Intelligence 8, 559–574 (2026). [62] B ̈ack, T. Evolutionary algorithms in theory and practice: evolution strategies, evolutionary programming, genetic algorithms (Oxford university press, 1996). [63] Brest, J., Greiner, S., Boskovic, B., Mernik, M. & Zumer, V. Self-adapting control parameters in differential evolution: A comparative study on numerical benchmark problems. IEEE transactions on evolutionary computation 10, 646– 657 (2006). [64] Gupta, A., Ong, Y.-S. & Feng, L. Multifactorial evolution: Toward evolutionary multitasking. IEEE Transactions on Evolutionary Computation 20, 343–357 (2015). [65] Wang, C., Liu, J., Wu, K. & Wu, Z. Solving multitask optimization problems with adaptive knowledge transfer via anomaly detection. IEEE Transactions on Evolutionary Computation 26, 304–318 (2021). [66] Pan, S. J. & Yang, Q. A survey on transfer learning. IEEE Transactions on knowledge and data engineering 22, 1345–1359 (2009). [67] Finn, C., Abbeel, P. & Levine, S.Model-agnostic meta-learning for fast adaptation of deep networks. arXiv preprint arXiv:1703.03400 (2017). [68] Yosinski, J., Clune, J., Bengio, Y. & Lipson, H. How transferable are features in deep neural networks? Advances in neural information processing systems 27 (2014). [69] Ayg ̈un, E. et al. An ai system to help scientists write expert-level empirical software. Nature 1–3 (2026). [70] Schmidhuber, J.G ̈odel machines: self-referential universal problem solvers making provably optimal self-improvements. arXiv preprint cs/0309048 (2003). [71] Stanley, K. O. & Miikkulainen, R. Evolving neural networks through augmenting topologies. Evolutionary computation 10, 99–127 (2002). [72] Watson, J. L. et al. De novo design of protein structure and function with rfdiffusion. Nature 620, 1089–1100 (2023). [73] Dauparas, J. et al. Robust deep learning–based protein sequence design using proteinmpnn. Science 378, 49–56 (2022). [74] Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O. & Dahl, G. E. Neural message passing for quantum chemistry. International Conference on Machine Learning 1263–1272 (2017). 17 [75] Sch ̈utt, K. T., Sauceda, H. E., Kindermans, P.-J., Tkatchenko, A. & M ̈uller, K.- R. Schnet–a deep learning architecture for molecules and materials. The Journal of chemical physics 148 (2018). [76] Batatia, I., Kovacs, D. P., Simm, G., Ortner, C. & Cs ́anyi, G. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields. Advances in neural information processing systems 35, 11423–11436 (2022). [77] Yao, S. et al. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems 36, 11809–11822 (2023). [78] Schick, T. et al. Toolformer: Language models can teach themselves to use tools. Advances in neural information processing systems 36, 68539–68551 (2023). [79] Madaan, A. et al. Self-refine: Iterative refinement with self-feedback. Advances in neural information processing systems 36, 46534–46594 (2023). [80] Deb, K., Pratap, A., Agarwal, S. & Meyarivan, T. A fast and elitist multiobjec- tive genetic algorithm: Nsga-i. IEEE transactions on evolutionary computation 6, 182–197 (2002). [81] Zhang, Q. & Li, H. Moea/d: A multiobjective evolutionary algorithm based on decomposition. IEEE Transactions on evolutionary computation 11, 712–731 (2007). [82] Gelbart, M. A., Snoek, J. & Adams, R. P. Bayesian optimization with unknown constraints. arXiv preprint arXiv:1403.5607 (2014). [83] Kandasamy, K., Dasarathy, G., Schneider, J. & Poczos, B.Multi-fidelity bayesian optimisation with continuous approximations.arXiv preprint arXiv:1703.06240 (2017). [84] Stokes, J. M. et al. A deep learning approach to antibiotic discovery. Cell 180, 688–702 (2020). [85] H ̈ase, F. et al. Olympus: a benchmarking framework for noisy optimization and experiment planning. Machine Learning: Science and Technology 2, 035021 (2021). [86] Real, E., Aggarwal, A., Huang, Y. & Le, Q. V. Regularized evolution for image classifier architecture search. Proceedings of the aaai conference on artificial intelligence 33, 4780–4789 (2019). [87] Li, L., Jamieson, K., DeSalvo, G., Rostamizadeh, A. & Talwalkar, A. Hyperband: A novel bandit-based approach to hyperparameter optimization. Journal of machine learning research 18, 1–52 (2018). 18 [88] Ye, X., Li, J., Wang, P. & Suganthan, P. N. A comprehensive survey of adap- tive strategies in differential evolutionary algorithms. Swarm and Evolutionary Computation 98, 102081 (2025). [89] Shields, B. J. et al. Bayesian reaction optimization as a tool for chemical synthesis. Nature 590, 89–96 (2021). [90] Seeger, M. Gaussian processes for machine learning. International journal of neural systems 14, 69–106 (2004). [91] Snoek, J., Larochelle, H. & Adams, R. P. Practical bayesian optimization of machine learning algorithms. Advances in neural information processing systems 25 (2012). [92] Shahriari, B., Swersky, K., Wang, Z., Adams, R. P. & De Freitas, N. Taking the human out of the loop: A review of bayesian optimization. Proceedings of the IEEE 104, 148–175 (2015). [93] Xie, T. & Grossman, J. C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Physical review letters 120, 145301 (2018). [94] Batzner, S. et al. E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications 13, 2453 (2022). [95] Ward, L., Agrawal, A., Choudhary, A. & Wolverton, C. A general-purpose machine learning framework for predicting properties of inorganic materials. npj Computational Materials 2, 16028 (2016). [96] Abramson, J. et al. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature 630, 493–500 (2024). [97] Rives, A. et al. Biological structure and function emerge from scaling unsu- pervised learning to 250 million protein sequences. Proceedings of the national academy of sciences 118, e2016239118 (2021). [98] Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123–1130 (2023). [99] Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373, 871–876 (2021). [100] Liang, X. et al. Sage: Self-evolving agents with reflective and memory-augmented abilities. Neurocomputing 647, 130470 (2025). [101] Krenn, M. et al. On scientific understanding with artificial intelligence. Nature Reviews Physics 4, 761–769 (2022). 19