PaperlayerView exact revision on arXiv

Source-linked AI summary

From AI for Science to Agentic Science: A Survey on Autonomous Scientific Discovery

Jiaqi Wei, Yuejin Yang, Xiang Zhang, Yuhan Chen, Xiang Zhuang, Zhangyang Gao, Dongzhan Zhou, Guangshuai Wang, Zhiqiang Gao, Juntai Cao, Zijie Qiu, Ming Hu, Chenglong Ma, Shixiang Tang, Junjun He, Chunfeng Song, Xuming He, Qiang Zhang, Chenyu You, Shuangjia Zheng, Ning Ding, Wanli Ouyang, Nanqing Dong, Yu Cheng, Siqi Sun, Lei Bai, Bowen Zhou

arXiv:2508.14111v2cs.LG

TL;DR

Autonomous scientific discovery lacks an integrated perspective spanning processes, autonomy, mechanisms, and domains. This survey unifies these perspectives in a framework connecting capabilities, processes, and domain realizations, establishing Agentic Science as a structured paradigm for AI-driven research.

  • Problem

    Prior surveys examine scientific discovery processes, autonomy, or architectures in isolation, leaving these perspectives insufficiently integrated across domains.

    Source: §1 · p. 5

  • Method

    The survey develops a comprehensive framework connecting foundational capabilities, core processes, and domain realizations across life sciences, chemistry, materials science, and physics.

    Source: §1 · p. 5

  • Results

    The survey provides a domain-oriented synthesis of autonomous scientific discovery and positions Agentic Science as a structured research paradigm.

    Source: §1 · p. 5, §11 · p. 49

  • Takeaways & Limitations

    Agentic Science offers a unified lens for understanding autonomous scientific discovery while highlighting technical, ethical, and philosophical challenges to trustworthy progress.

    Source: §11 · p. 49

    Current agentic systems remain poorly equipped for long-term causal reasoning across experiments conducted months or years apart.

    Source: memory mechanisms · p. 25

Abstract

from arXiv · show

Artificial intelligence (AI) is reshaping scientific discovery, evolving from specialized computational tools into autonomous research partners. We position Agentic Science as a pivotal stage within the broader AI for Science paradigm, where AI systems progress from partial assistance to full scientific agency. Enabled by large language models (LLMs), multimodal systems, and integrated research platforms, agentic AI shows capabilities in hypothesis generation, experimental design, execution, analysis, and iterative refinement -- behaviors once regarded as uniquely human. This survey provides a domain-oriented review of autonomous scientific discovery across life sciences, chemistry, materials science, and physics. We unify three previously fragmented perspectives -- process-oriented, autonomy-oriented, and mechanism-oriented -- through a comprehensive framework that connects foundational capabilities, core processes, and domain-specific realizations. Building on this framework, we (i) trace the evolution of AI for Science, (ii) identify five core capabilities underpinning scientific agency, (iii) model discovery as a dynamic four-stage workflow, (iv) review applications across the above domains, and (v) synthesize key challenges and future opportunities. This work establishes a domain-oriented synthesis of autonomous scientific discovery and positions Agentic Science as a structured paradigm for advancing AI-driven research.

1. Introduction

The paper frames Agentic Science as a pivotal AI for Science stage in which autonomous agents conduct iterative scientific discovery, and proposes an integrated framework spanning capabilities, processes, and domain applications. It reviews this paradigm across life sciences, chemistry, materials science, and physics while synthesizing its challenges and future opportunities.

Evolution of AI for Science

  • Agentic Science marks the transition from AI systems serving targeted computational tasks to autonomous agents that formulate hypotheses, design and execute experiments, interpret results, and iteratively refine theories [288].It aligns primarily with Level 3, Full Agentic Discovery, while building on Level 2, Partial Agentic Discovery (Figure 1).Source: §1 · p. 4, Figure 1 · p. 4

Unified Framework

  • The framework integrates process-oriented, autonomy-oriented, and mechanism-oriented perspectives that prior surveys have treated separately.It connects foundational capabilities, core processes, and domain realizations within autonomous scientific discovery.Source: §1 · pp. 4–5

Core Capabilities

  • Scientific agents rely on five foundational capabilities: reasoning and planning, tool integration, memory mechanisms, multi-agent collaboration, and optimization and evolution.The paper reviews state-of-the-art implementations and domain-specific challenges for each capability [219] [35] [46].Source: §1 · p. 5

Dynamic Workflow

  • Agentic Science models discovery as a dynamic four-stage workflow: observation and hypothesis generation, experimental planning and execution, data and result analysis, and synthesis, validation, and evolution.Agents can flexibly and dynamically combine these stages to address complex scientific problems [97] [96].Source: §1 · p. 5

Applications and Future Directions

  • The survey reviews agentic systems across life sciences, chemistry, materials, and physics, covering more than a dozen subfields from drug discovery to materials design [147].It also synthesizes technical, ethical, and philosophical challenges, including reproducibility, validation of novel discoveries, and human-agent collaboration.Source: §1 · p. 5

2. The Evolution of AI for Science: From Tools to Autonomous Partners

AI for Science is evolving from narrowly scoped computational tools toward autonomous, end-to-end scientific inquiry. This progression advances through distinct levels of autonomy and capability, culminating in Agentic Science as an autonomous scientific-partner paradigm.

Evolution of AI for Science

  • AI for Science progresses from specialized tools to autonomous scientific partners through increasing levels of autonomy and capability.The section formally delineates these levels and frames Agentic Science as their culmination.Source: §2 · p. 5

Research framework

  • Figure 2 integrates foundational capabilities, core processes, and research levels for autonomous scientific discovery across four scientific domains.The framework spans life sciences, chemistry, materials science, and physics.Source: Figure 2 · p. 6

2.1. The Evolution of AI for Science

AI for Science evolves from specialized, human-directed computational tools to partially autonomous assistants and then autonomous scientific partners, with a prospective fourth stage focused on inventing new scientific frameworks. This progression shifts AI from prediction and execution toward independent hypothesis generation, experimentation, refinement, and scientific creation.

2.1.1. Level 1: AI as a Computational Oracle (Expert Tools)

  • Level 1 treats AI as a Computational Oracle: specialized models produce predictions or generated outputs for defined tasks, while humans retain control of hypotheses, experiments, execution, and interpretation.These models function as static function approximators without capacity for subsequent independent action, despite advances across life sciences, chemistry, materials science, physics, and astronomy.Source: §2.1.1 · pp. 6–7

2.1.2. Level 2: AI as an Automated Research Assistant (Partial Agentic Discovery)

  • Level 2 introduces Automated Research Assistants that integrate tools and execute sequences for human-defined subgoals, but scientific direction and the initial hypothesis remain human-provided.After completing the predefined subgoal, control returns to the human scientist; applications include bioinformatics, experimental design, reaction optimization, automated experimentation, simulations, power-electronics design, and spacecraft control.Source: §2.1.2 · p. 7

2.1.3. Level 3: AI as an Autonomous Scientific Partner (Full Agentic Discovery)

  • Level 3 defines full Agentic Science: AI independently observes domains, formulates hypotheses, designs and executes experiments, analyzes results, and iteratively refines its strategy with minimal human intervention.The human becomes a high-level strategist and validator, while the agent operates in an open-ended discovery loop that updates knowledge, evidence, and hypotheses.Source: §2.1.3 · p. 7, Figure 3 · p. 8

2.1.4. Level 4: AI as a Generative Architect (Future Prospect)

  • Level 4 envisions a Generative Architect that invents scientific instruments, methodologies, conceptual frameworks, and interdisciplinary principles rather than operating only within existing scientific paradigms.Its prospective endpoint is a Global Cooperation Research Agent whose achievement could be evaluated by a “Nobel-Turing Test.”Source: §2.1.4 · p. 8

2.2. The Human Scientist’s Evolving Role

Agentic AI shifts scientists from executing research tasks to strategically defining goals, supervising autonomous work, and integrating results into reliable scientific narratives. This transition also redistributes innovation across coordinated human-agent systems while requiring new skills, oversight practices, and transparency standards.

The Human Scientist’s Evolving Role

  • Agentic AI moves scientists from executors to strategists who define research goals, maintain ethical and reliable methods, and integrate results into coherent narratives.Source: §2.2 · p. 9
  • Scientists guide autonomous discovery by specifying goals G = {g1, . . . , gm}, embedding safety and ethical limits, validating results, and correcting deviations from intended objectives.Source: §2.2 · p. 9
  • Effective oversight requires context-rich agent prompting, toolset management, calibrated trust, and deeper scrutiny when outputs require verification.Research culture may develop human-on-the-loop laboratories, human-agent evaluation by funders, and journal requirements for complete agent traces, code, and tool use.Source: §2.2 · p. 9
  • Innovation becomes distributed across human-agent systems, with agents pursuing subgoals Gsub ⊂G, exploring large problem spaces, detecting beyond-human patterns, and connecting cross-disciplinary ideas [91].This distribution can speed discovery and expand the questions science can ask and answer [91].Source: §2.2 · p. 9

2.3. Agentic Science: The Focus of This Survey

This survey focuses on Agentic Science, where AI systems act as goal-directed scientific agents with autonomous reasoning, planning, and action. It organizes the field through foundational capabilities, core discovery processes, and autonomous scientific discovery across major research domains.

2.3. Agentic Science: The Focus of This Survey

  • The survey uses a three-level framework linking Five Foundational Capabilities to Four Core Processes and then to Autonomous Scientific Discovery across life sciences, chemistry, materials science, and physics.The capabilities form the cognitive core of scientific agents, enabling the agentic discovery loop and domain-specific advances.Source: §2.3 · p. 9
  • Agentic Science encompasses task-level and goal-level autonomy, ranging from automating specific research tasks to independently pursuing high-level scientific objectives [288] [106] [278].Its defining commonality is agency: purposeful, independent action within an environment to achieve a goal.Source: §2.3 · p. 9
  • Its central theme is the development of agentic capabilities spanning planning and reasoning, tool use, memory, collaboration, and evolution [138] [100] [102] [347] [320] [218] [127] [151].The survey examines progress and challenges across these capabilities as agentic AI develops toward reshaping scientific discovery.Source: §2.3 · p. 9

2.4. Modern Scientific Large Language Models for Agentic Science

Scientific Large Language Models provide the technological foundation for Agentic Science, with general-purpose reasoning and tool-use models enabling higher autonomy levels than specialized computational tools. Their development and domain-specific adaptations support scientific workflows across life sciences, chemistry, materials science, physics, and astronomy.

Foundations

  • General-purpose Sci-LLMs with advanced reasoning and tool use underpin Level 2 and 3 agentic systems, whereas specialized models mainly function as Level 1 computational oracles.The section frames Sci-LLMs as the practical technological bedrock for the shift from AI tools to autonomous scientific partners.Source: §2.4 · p. 10

Development Strategies

  • Sci-LLMs are developed through science-focused fine-tuning, scientific pre-training, and test-time reasoning enhancements using Chain-of-Thought, reinforcement learning, and Mixture-of-Experts architectures.Fine-tuning generally produces Level 1 expert tools, while large-scale pre-training and enhanced reasoning enable Levels 2 and 3.Source: §2.4 · p. 10

Life Sciences

  • Life-science Sci-LLMs support sequence representation, cross-omics analysis, molecular design, medical dialogue, and multimodal diagnostic reasoning across biological and healthcare data.Applications include DNA, RNA, and protein modeling; molecular property prediction; single-cell analysis; drug discovery; medical imaging; and visual question answering.Source: §2.4 · p. 10

Chemistry

  • Chemistry Sci-LLMs bridge molecular structures, reactions, and three-dimensional conformations for property prediction, retrosynthesis, and structure-based drug design.These models connect text, SMILES and other two-dimensional representations, reaction data, and 3D geometries.Source: §2.4 · p. 10

Materials Science

  • Materials-science models predict properties, generate molecules and crystal structures, and address synthesis, reaction, and mechanics tasks using domain knowledge and retrieval augmentation.Architectures include representation encoders, decoder-only generators, and specialized models integrating retrieval-augmented generation with knowledge graphs.Source: §2.4 · p. 11

Physics

  • Physics Sci-LLMs integrate language processing with physics engines and visual modules to estimate parameters, learn PDE solution operators, and retrieve high-energy-physics knowledge.They are evolving into interactive tools for scientific workflows rather than serving only as text-based systems.Source: §2.4 · p. 11

Astronomy

  • Astronomy models adapt general architectures through continual literature pre-training and task-specific fine-tuning, with multimodal variants interpreting astronomical images such as galaxy observations.These adaptations improve domain-specific astronomical text understanding, generation, and visual description.Source: §2.4 · p. 11

3. Scientific Agents: Core Abilities and Challenges

Scientific agents require five interconnected abilities—planning and reasoning, tool integration, memory, collaboration, and optimization/evolution—to manage long-term, iterative, empirically grounded discovery workflows. Each ability enables scientific agency while introducing challenges involving reliability, multimodal evidence, reproducibility, uncertainty, resource constraints, and safety.

3. Scientific Agents: Core Abilities and Challenges

  • Scientific agents combine five foundational pillars: planning and reasoning, external tools, memory, collaboration, and continuous optimization and evolution.Together, these capabilities support autonomous navigation of long-term, iterative, empirically grounded research workflows rather than discrete short-horizon tasks.Source: §3 · p. 11

3.1. Planning and Reasoning Engines

  • Planning engines translate scientific goals into executable actions and adapt through search, feedback, reflection, and reinforcement learning.Scientific planning must handle vast search spaces, noisy multimodal feedback, long horizons, causal understanding, strict verifiability, and error accumulation.Source: §3.1 · pp. 11, 13, 14

3.2. Tool Use and Integration

  • Tool integration extends agents from general computation and information retrieval to domain-specific analysis and experimental or simulation platforms for hypothesis testing.Scientific tool use additionally requires precise parameterization, provenance and reproducibility, interoperable workflows, adaptation to changing interfaces, and cost-aware resource management.Source: §3.2 · pp. 14–15

3.3. Memory Mechanisms

  • Memory supports iterative task execution and knowledge retrieval by preserving experience, reusable skills, literature, databases, and structured scientific concepts.Scientific memory must validate aging knowledge, integrate heterogeneous multimodal data, and preserve long-term causally linked histories with sufficient fidelity for reproducibility.Source: §3.3 · pp. 15–17

3.4. Collaboration between Agents

  • Multi-agent collaboration improves robustness and creativity through hierarchical workflows, deliberative refinement, and adaptive topologies.Scientific collaboration must produce testable and verifiable outcomes, preserve epistemic diversity and uncertainty, avoid amplified hallucinations and premature consensus, and coordinate multimodal data, software, and instruments.Source: §3.4 · pp. 17–19

3.5. Optimization and Evolution

  • Optimization and evolution improve scientific strategies, knowledge structures, and architectures through self-refinement, self-learning, tool feedback, and population-based co-evolution.Scientific improvement is constrained by expensive evaluations, sparse rewards, physical validity, safety protocols, and the requirement for reproducible, verifiable knowledge.Source: §3.5 · pp. 19–21

4. Agentic Science: Dynamic Workflow and Challenges

Agentic Science frames discovery as a self-improving, closed-loop workflow spanning hypothesis generation, experimentation, analysis, and synthesis, with execution order adapting to objectives, context, and results. Fully autonomous pipelines extend this loop through multi-agent collaboration, self-evolution, human feedback, and specialized evaluation environments.

4. Agentic Science: Dynamic Workflow and Challenges

  • Agentic Science comprises four dynamically ordered stages: observation and hypothesis generation, experimental planning and execution, result analysis, and synthesis, validation, and evolution.Not every system requires every stage, and agents may adjust execution order according to objectives, context, and ongoing results.Source: §4 · p. 21

4.1. Observation and Hypothesis Generation

  • OriGene identified GPR160 and ARG2 as novel cancer targets validated in patient-derived systems, while Robin [97] and CellVoyager surfaced additional therapeutic and aging-related insights.These systems combine multimodal data, autonomous background research, inference, and transcriptomic reanalysis to generate testable hypotheses.Source: §4.1 · p. 23

4.2. Experimental Planning and Execution

  • Coscientist [30] and ORGANA [67] demonstrated robotic experimental autonomy, with ORGANA completing a 19-step synthesis while reducing human workload by over 80%.The Virtual Lab also designed 92 SARS-CoV-2 nanobodies computationally, two of which showed strong binding in later empirical tests.Source: §4.2 · p. 24

4.3. Data and Result Analysis

  • Agentic analysis combines multimodal extraction, structured reasoning, and memory-based Bayesian updating, but must distinguish signal from noise across ambiguous scientific data types.Challenges include confirmation bias and integrating text, tables, genomic sequences, and imagery into coherent conclusions.Source: §4.3 · p. 24
  • Robin [97], PROTEUS [71], SpatialAgent, and LLM-RDF demonstrate autonomous interpretation that extracts mechanisms, analyzes large datasets, and informs subsequent experimental steps.SpatialAgent achieved expert-level performance on spatial biology datasets containing over two million single-cell measurements.Source: §4.3 · p. 24

4.4. Synthesis, Validation, and Evolution

  • Synthesis and validation use evidence critique, deliberative multi-agent review, reproducibility checks, and adaptive policy refinement to improve future research decisions.Long-horizon causal reasoning remains difficult because agents must preserve faithful, causally linked histories across experiments separated by months or years.Source: memory mechanisms · p. 25

memory mechanisms.

  • Sparks [96], OriGene, CellAgent [367], and MOOSE-Chem3 [210] demonstrate iterative self-correction, feedback assimilation, recursive evaluation, and experiment-guided candidate prioritization.These systems close the loop between evaluation and exploration by learning from prior outcomes to guide subsequent hypotheses or experiments.Source: memory mechanisms · p. 25

4.5. Fully Autonomous Research Pipeline

  • Fully autonomous pipelines span idea generation through reporting, with systems such as The AI Scientist [219], Coscientist, and AlphaEvolve [250] demonstrating end-to-end research across scientific domains.Multi-agent collaboration, self-evolution, human feedback, and environments such as DiscoveryWorld [146] support broader capabilities, improved performance, alignment, and repeatable evaluation.Source: §4.5 · pp. 25–27

5. Agentic Life Sciences Research

Agentic AI is transforming life-sciences research by coordinating specialized agents across data analysis, hypothesis generation, experimental design, execution, and interpretation. Applications span general biomedical frameworks, multi-omics, protein engineering, and drug discovery, with systems increasingly achieving autonomous, validated scientific outcomes.

5. Agentic Life Sciences Research

  • Across life sciences, agentic systems address high-dimensional datasets and multi-step workflows through specialized collaborating agents that automate analysis, experimentation, interpretation, and discovery.The surveyed scope includes genomics, proteomics, drug discovery, and protein engineering, with representative tasks organized by primary application domain in Tables 8 and 9.Source: §5 · pp. 27–28, Figure 7 · p. 28

5.1. General Frameworks and Methodologies

  • General biomedical frameworks combine modular multi-agent collaboration, self-evolving tools, dynamic planning, quality control, and scientific principles to make autonomous research assistants more adaptable.STELLA’s benchmark accuracy improves systematically with trials, BioResearcher achieved a 63.07% average execution success rate across eight unmet objectives, and PiFlow increased discovery efficiency by 73.55% while improving solution quality by 94.06%.Source: §5.1 · pp. 28–29

5.2. Genomics, Transcriptomics, and Multi-Omics Analysis

  • Multi-omics agents automate complex workflows from natural-language analysis and visualization to iterative experiment design and autonomous hypothesis generation across single-cell, spatial, transcriptomic, proteomic, and mass-spectrometry data.CellAgent refined tools and hyperparameters without human intervention, CellVoyager discovered a previously unreported link between transcriptional noise and brain aging, and PerTurboAgent outperformed active-learning strategies for impactful perturbations.Source: §5.2 · pp. 30–31

5.3. Protein Science and Engineering

  • Protein-engineering agents integrate retrieval, structure analysis, simulation, sequence generation, prediction, and reflection to design proteins and autonomously discover verifiable design principles.ProtAgents designed proteins with targeted mechanical properties, while Sparks uncovered a length-dependent mechanical crossover in peptide unfolding force and a chain-length/secondary-structure stability map.Source: §5.3 · p. 31

5.4. Drug and Therapeutic Discovery

  • Drug-discovery agents now support end-to-end pipelines and validated therapeutic discovery, including novel targets, repurposing candidates, and experimentally improved nanobody binders.LIDDiA generated molecules meeting pharmaceutical criteria for over 70% of 30 clinically relevant targets; OriGene nominated GPR160 and ARG2, Robin identified ripasudil for dAMD, and the Virtual Lab experimentally validated two improved binders from 92 candidates.Source: §5.4 · pp. 32–33

6. Agentic Chemistry Research

Agentic AI is transforming chemistry by integrating LLMs, specialized tools, and robotic platforms across hypothesis generation, experimentation, analysis, and refinement. The surveyed systems span general agent architectures, autonomous synthesis, generative molecular design, and computational chemistry workflows.

6. Agentic Chemistry Research

  • Agentic chemistry systems integrate LLMs with chemical tools and robotic platforms to automate hypothesis generation, experimental execution, analysis, material discovery, and reaction optimization.The survey organizes these systems by their primary functions and highlights associated challenges, frameworks, and achievements in Tables 10 and 11.Source: §6 · p. 33

6.1. General Frameworks and Methodologies

  • General-purpose chemistry agents address tool integration, reasoning, hypothesis generation, and literature comprehension through multi-agent architectures, tool-enhanced planning, experiment-guided ranking, and self-updating knowledge libraries.MOOSE-Chem rediscovered the core innovations of 51 recent high-impact papers, while Chemagent achieved performance gains of up to 46% on SciBench; tool benefits vary across task types.Source: §6.1 · pp. 33–35

6.2. Organic Synthesis and Reaction Optimization

  • Autonomous synthesis systems now coordinate reaction design, robotic execution, analysis, and scale-up, including closed-loop discovery across five laboratories that produced 21 new state-of-the-art organic laser emitters.Coscientist optimized palladium-catalyzed cross-couplings, while LLM-RDF coordinated multi-agent synthesis development and ORGANA reduced physical demand by over 50% and saved researchers an average of 80% of their time.Source: §6.2 · pp. 35–36

6.3. Generative Chemistry and Molecular Design

  • Generative chemistry agents combine language models, structure generation, quantum validation, and synthetic-feasibility assessment to create novel molecules and materials, including five experimentally synthesized AI-dreamt MOFs.MOFGen generated hundreds of thousands of novel MOF structures, while other systems use self-reflection, quantum feedback, or reaction-constrained reinforcement learning to guide practical candidate design.Source: §6.3 · pp. 36–37

6.4. Computational and Quantum Chemistry

  • Agentic computational chemistry assistants translate natural-language requests into executable simulations, using adaptive memory, tool selection, workflow planning, and validation to support quantum and atomistic chemistry.El Agente Q achieved an average success rate of over 87% on benchmark tasks, while xChemAgents reduced mean absolute error by up to 22% over baselines with interpretable explanations.Source: §6.4 · p. 37

7. Agentic Materials Science Research

Agentic AI in materials science is organized around general discovery platforms, novel-material design, and automated simulation and characterization. These systems integrate data, planning, scientific reasoning, human collaboration, and iterative validation across complex workflows.

7. Agentic Materials Science Research

  • Materials-science agentic research spans general discovery platforms, novel-material design, and automated simulation and characterization workflows.The survey identifies these three application areas as the section’s organizing structure.Source: §7 · p. 37

7.1. General Methodologies and Discovery Platforms

  • General platforms address siloed data, unreliable LLM knowledge, workflow planning, and human collaboration through retrieval, tool use, hypothesis evaluation, and multimodal coordination.LLaMP [58] grounds agents in materials databases and simulations; HoneyComb [423] uses MatSciKB and a tool hub; MatPilot [248] supports closed-loop human collaboration; MAPPS [444] automates workflow planning and reports a five-fold improvement in stable, novel crystal-structure generation; a multicrossmodal framework [25] unifies images, text, tables, and videos.Source: §7.1 · pp. 38–39

7.2. Design and Discovery of Novel Materials

  • Multi-agent systems navigate vast materials-design spaces by generating hypotheses, applying physical or computational feedback, and discovering new biocomposites, alloys, topological phases, metamaterials, and user-specified structures.SciAgents, AtomAgents, and GNN-assisted alloy discovery [92] target biocomposites and alloys; TopoMAS [413] identified SrSbO3; CrossMatAgent [326], a photonic inverse-design framework [220], and metaAgent address metamaterials and electromagnetic tasks; dZiner [8], LLMatDesign [149], PriM [175], and PiFlow [266] support inverse and principle-guided design.Source: §7.2 · p. 40

7.3. Automated Simulation and Characterization

  • Agentic systems lower barriers to computational and experimental materials workflows by interpreting natural-language tasks, decomposing simulations, correcting errors, and connecting microscopy with atomistic modeling.Foam-Agent [407] achieved 83.6% success on CFD benchmark tasks; ChemGraph [262] supports geometry optimization and thermochemistry; MechAgents [247] self-corrects finite-element analyses; AILA [232] found multi-agent architectures outperform single agents in AFM workflows; AutoMat [391] reconstructs crystal structures and predicts properties from microscopy images.Source: §7.3 · pp. 40–41

8. Agentic Physics and Astronomy Research

Agentic AI is accelerating physics and astronomy research by coordinating software, data analysis, hypothesis testing, experiments, and broader research workflows. Across general methodologies, astronomy, computational mechanics, and quantum computing, these systems demonstrate increasingly autonomous and reliable scientific operation.

8.1. General Frameworks and Methodologies

  • Agentic frameworks improve scientific reasoning and system design through iterative refinement, physics-informed planning, autonomous spacecraft control, tool orchestration, and experiment execution.MoRA improved open-source LLM physics-benchmark accuracy by up to 16% [145], while LP-COMDA reduced power-converter design error by 63.2% and was over 33 times faster than human-led processes [206]. Accelerator Assistant reduced synchrotron experiment preparation time by two orders of magnitude [121].Source: §8.1 · p. 43

8.2. Astronomy and Cosmology

  • Astronomy agents automate telescope operations, interpret JWST observations, and conduct cosmological research from planning and coding through analysis and publication.StarWhisper automates NGSS observation workflows and follow-up proposals, while mephisto achieved near-human JWST analysis and identified novel “Little Red Dot” galaxy populations [311]. AI Cosmologist autonomously performs machine-learning research through paper writing [241], and an autogen/ag2 system supports cosmological parameter analysis using RAG and local execution [178].Source: §8.2 · pp. 43–44

8.3. Computational Mechanics and Fluid Dynamics

  • OpenFOAMGPT 2.0 achieved 100% success and reproducibility across over 450 test cases by orchestrating specialized agents for end-to-end CFD simulations [78].Related systems reached over 99.0% accuracy for structural beam analysis [205], used planner-coder-critic collaboration for elasticity problems [247], and automated research workflows from idea generation to scientific publication.Source: §8.3 · p. 44

8.4. Quantum Computing

  • The k-agents framework autonomously planned and executed superconducting-processor experiments for hours, producing and characterizing entangled quantum states with proficiency matching human scientists.Its execution agents encode laboratory knowledge as state machines and use analyzed results to drive closed-loop state transitions.Source: §8.4 · p. 45

9. Challenges in Agentic Science

Agentic Science introduces foundational challenges involving reproducibility, novelty validation, interpretability, and ethical governance. Addressing these concerns is essential for safely and credibly integrating autonomous AI into scientific research.

9. Challenges in Agentic Science

  • Agentic Science raises foundational and ethical challenges that affect how scientific knowledge is generated, validated, trusted, and safely integrated into the natural sciences.These concerns extend beyond established LLM limitations such as hallucination, inefficient knowledge updating, and catastrophic forgetting [165].Source: §9 · p. 45

9.1. Agentic Reproducibility and Reliability

  • Agentic discovery trajectories are difficult to reproduce because stochastic reasoning, planning and execution failures, system instability, and prompt sensitivity produce divergent outcomes.Execution accuracy on state-of-the-art benchmarks can be as low as 39% [364], while catastrophic forgetting undermines knowledge continuity and model updates can prevent earlier findings from being reproduced [165]. Formal logging of agent states, policies, justifications, and environmental conditions is therefore required.Source: §9.1 · pp. 45–46

9.2. Validation of Novelty

  • Novel hypotheses are difficult to validate because apparent conceptual breakthroughs may instead reflect interpolation, repetition, hallucination, or fabricated findings, data, and references.Auditing reasoning lineage, improving interpretability, and developing objective scalable evaluation frameworks remain necessary because current assessment relies on costly and subjective expert judgment.Source: §9.2 · p. 46

9.3. Transparency in Scientific Reasoning

  • Opaque AI reasoning undermines scientific validation, trust, and knowledge integration, requiring interpretable-by-design agents supported by structured internal logs and clear documentation [28] [23] [22].Interpretability enables researchers to audit whether conclusions follow sound logic and to critically evaluate and build upon machine-generated knowledge.Source: §9.3 · p. 46

9.4. Ethical and Societal Dimensions

  • Autonomous discovery agents create accountability, dual-use, labor, educational, governance, integrity, and bias risks that require attribution, traceability, normative constraints, audit trails, oversight, and preserved human agency.Hazardous discoveries and adversarial attacks raise misuse concerns [119] [451], while over-reliance may erode skills [381], overwhelm peer review [219], and intensify funding inequalities [381].Source: §9.4 · p. 47

10. Future Outlook of Agentic Science

The future of Agentic Science extends beyond workflow automation toward autonomous invention, scalable interdisciplinary synthesis, global agent cooperation, and fully autonomous scientific discovery. These directions could transform how scientific knowledge is invented, justified, and disseminated.

10. Future Outlook of Agentic Science

  • Agentic Science’s future requires bridging conceptual, technical, and ethical gaps while advancing toward computational epistemology and four prospective frontiers.The envisioned paradigm moves beyond incremental automation toward artificial agents participating in scientific knowledge production and dissemination.Source: §10 · p. 47

10.1. From Automation to Autonomous Invention

  • Autonomous invention would transform agents from tool-users into tool-creators that identify methodological limits and propose novel instruments or conceptual frameworks.Examples include inventing imaging modalities for inaccessible subcellular processes or mathematical abstractions for emergent behaviors.Source: §10.1 · pp. 47–48

10.2. Interdisciplinary Synthesis at Scale

  • Multimodal agents could enable interdisciplinary synthesis at scale by mapping latent analogies across fields and discovering unifying principles beyond information retrieval.Illustrative connections include topological quantum field theory with deep learning and ecological dynamics with economic systems.Source: §10.2 · p. 48

10.3. The Global Cooperation Research Agent

  • A decentralized global network of specialized agents could share data, conduct peer review, refine hypotheses, and coordinate experiments on grand challenges [70].Realization would require federated protocols, secure multi-agent reasoning against recursive attack propagation [449], and proof-of-thought cryptographic trails for accountability.Source: §10.3 · p. 48

10.4. The Nobel-Turing Test

  • The Nobel-Turing Test asks whether an autonomous agent or human-agent team can produce a Nobel-worthy discovery by completing the full scientific cycle.This requires identifying a foundational gap, generating a non-obvious testable hypothesis, designing novel experiments, and interpreting results to trigger a paradigmatic shift [45].Source: §10.4 · p. 49

11. Conclusion

Agentic Science is framed as a transformative stage of AI for Science, unifying capabilities, processes, and domain realizations into a paradigm for autonomous scientific discovery. It broadens research potential while requiring technical, ethical, and philosophical safeguards, positioning AI as a co-evolving partner that augments human inquiry.

Conclusion

  • Agentic Science advances AI for Science from computational assistance toward autonomous research partnerships involving reasoning, experimentation, and iterative discovery.The paradigm spans life sciences, chemistry, materials science, and physics through a unified framework connecting foundational capabilities, core processes, and domain realizations.Source: §11 · p. 49
  • The paradigm promises broad applicability and accelerated scientific discovery, but requires technical, ethical, and philosophical challenges to be addressed for trustworthy and impactful progress.It is presented as augmenting rather than replacing human inquiry, supporting scientific creativity while reshaping the future of research.Source: §11 · p. 49

Related papers

Loading 2508.14111v2…