Towards Scientific Intelligence: A Survey of LLM-based Scientific Agents

• LLM-based scientific agents automate hypothesis generation, experimental design, and data analysis while handling domain-specific knowledge, diverse action spaces, and heterogeneous scientific data.Their specialization supports interdisciplinary discovery and accelerates scientific research beyond the broad applications of general-purpose LLM agents.Source: §1 · p. 1
• The survey’s mechanism-oriented taxonomy centers on four architectural building blocks—planner, memory, action space, and verifier—to explain trustworthy, reproducible, and scientifically valid autonomy.This design-focused perspective complements lifecycle- and role-based surveys by connecting high-level capabilities to underlying architectural and algorithmic principles.Source: §1 · pp. 1–2
• The survey provides a component-wise construction blueprint with multiple mechanism subtypes, mix-and-match assembly guidance, and an end-to-end cathode-design example.The example functions as a practical recipe for constructing fit-for-purpose scientific agents.Source: §1 · p. 2
• It synthesizes more than 120 representative papers and 40 domain benchmarks into fine-grained mechanism-level categories to help researchers locate transferable techniques and baselines.The related works are also grouped by domain to support rapid understanding of how to build scientific agents for particular fields.Source: §1 · p. 2
• Ethics, bias mitigation, and reproducibility are treated as intrinsic architectural and verification constraints rather than peripheral concerns.The survey also identifies interdisciplinary knowledge integration, dynamic adaptation, and standardized reproducibility protocols as key future directions.Source: §1 · p. 2

2 Architecture

• The architecture comprises Planner, Memory, Action Space, and Verifier components that iteratively decompose queries, execute tools, check intermediate results, and store verified outputs until validity and reproducibility standards are met.The Action Space may include APIs, simulators, laboratory instruments, search engines, or the LLM itself; multimodal perception is integrated into the Planner rather than separated into a perceptron module.Source: §2 · p. 4

2.1 Planner

• The Planner translates high-level research objectives into executable task sequences coordinating tool invocations, memory operations, and verification, while balancing granularity, dependencies, and adaptive replanning under scientific uncertainty.Scientific planning can cover hypothesis formulation, experimental design, data analysis, and validation; planners are categorized as prompt-native or learned according to their operating mechanisms.Source: §2.1 · p. 4

2.1.1 Prompt-Native Planners

• Prompt-native planners use prompts rather than parameter updates to construct interpretable, rapidly adaptable workflows, spanning instructional/schema-driven, context-augmented, deliberative/reflective, search-based, role-interactive, and programmatic subtypes.Some systems combine multiple prompt types, so the taxonomy identifies typical rather than necessarily exclusive categories.Source: §2.1.1 · p. 4, Figure 3 · p. 5
• Schema-driven planners embed procedural templates, response formats, tool inventories, domain guidelines, and task-specific personas into prompts, producing structurally consistent and reproducible plans but adapting poorly to unanticipated problems.Their performance depends critically on schema quality and completeness.Source: §2.1.1 · pp. 4, 6, 7
• Context-augmented planners retrieve literature, data, knowledge-graph entries, archives, or documentation and inject that evidence into prompts to guide prioritization, feasibility checks, and methodology selection.This grounding improves accuracy and reproducibility, but overly large or noisy contexts can dilute focus and cause inconsistency or factual drift.Source: §2.1.1 · pp. 7–8
• Deliberative, search-based, and role-interactive planners improve planning through distinct forms of refinement: critique-and-revision loops, evaluated exploration of multiple candidate plans, or collaborative dialogue among specialized agents.These approaches respectively depend on the LLM’s introspection, incur computational and scalability costs, or distribute planning across role-specific prompts and consensus protocols.Source: §2.1.1 · pp. 8–10
• Prompt-native planners coordinate specialized agents or generate executable DSL, code, and DAG artifacts, improving workflow precision and reproducibility while introducing coordination overhead or rigidity.Multi-agent systems use role-based collaboration and validation, whereas programmatic planners encode dependencies and control flow for direct downstream execution.Source: §2.1.1 · pp. 11–12

2.1.2 Learned Planners

• Learned planners internalize planning logic through supervised or domain training, reinforcement learning, and preference optimization, enabling domain-specific autonomy and potentially novel strategies beyond expert imitation.Domain-trained planners are stable in-distribution but struggle out of distribution and offer limited interpretability, while RL-based planners require careful rewards, substantial computation, and efficient evaluation.Source: §2.1.2 · pp. 12–14

2.1.3 Discussion

• Scientific-agent planners include sequential, hierarchical, graph, tree-search, and reactive designs, with hybrid architectures combining templates, reflection, collaboration, and learned components.Future systems must support open-ended discovery, multitimescale coordination, and adaptive replanning under uncertainty rather than remaining task-specific.Source: §2.1.3 · p. 14

2.2 Memory

• Memory provides the epistemic foundation for scientific agents by retaining findings, refining hypotheses, integrating external knowledge, avoiding redundant exploration, and building on prior discoveries.The survey categorizes memory as Historical Context, External Knowledge Base, and Intrinsic Knowledge, comparing these mechanisms in Table 1 and illustrating them in Figures 5 and 6.Source: §2.2 · p. 14

2.2.1 M1. Historical Context

• Historical context combines volatile session memory with persistent cross-session storage, allowing agents to accumulate interactions, outcomes, and reasoning steps for iterative, reproducible research.Hybrid memory balances immediate accessibility for real-time decisions against durable retention for cumulative learning across research workflows.Source: §2.2.1 · p. 14

M1. Historical Context

• Recent scientific-agent work includes MetaAgent and STELLA (Jin et al., 2025).Source: M1. Historical Context · p. 15
• Domain-focused examples include AccelMat (Kumbhar et al., 2025), BiomedRAG (Li et al., 2024c), and Agent Laboratory (Schmidgall et al., 2025).Source: M1. Historical Context · p. 15
• Additional systems address biology, chemistry, drugs, materials, geoscience, simulation, and general scientific research, including Coscientist, DrugAgent (Inoue et al., 2024), and ResearchAgent (Baek et al., 2024).Source: M1. Historical Context · p. 15

M3. Intrinsic Knowledge

• Historical context stores conversational turns, tool outputs, intermediate results, error traces, multi-agent communications, and tool states to support iterative refinement and self-optimization.Advanced systems additionally retain validated solutions, replay past trajectories, or restore state across sessions.Source: M3. Intrinsic Knowledge · p. 16

2.2.2 M2. External Knowledge Base

• External knowledge bases use query-driven retrieval, including semantic search, structured queries, RAG, knowledge graphs, APIs, and web browsing, to ground reasoning in current specialized information.They improve hypothesis generation, experimental design, data analysis, transparency, and factual reliability, but face latency, retrieval-quality, and context-dilution problems.Source: §2.2.2 · pp. 17, 19

2.2.3 M3. Intrinsic Knowledge

• Intrinsic knowledge encodes scientific concepts, terminology, procedures, and reasoning patterns in model parameters, enabling fluent domain reasoning without external queries.It is acquired through pre-training and task-specific adaptation, although most agents activate existing foundation-model knowledge through prompting and role specification.Source: §2.2.3 · pp. 19–20
• Intrinsic knowledge offers immediate, zero-latency scientific understanding and broad generalization, but suffers from stale information, opaque provenance, weak source attribution, hallucination risk, and costly updating.These limitations motivate combining parametric knowledge with external knowledge bases and historical context.Source: §2.2.3 · p. 20

2.2.4 Discussion

• Hybrid architectures combine historical context, external retrieval, and intrinsic knowledge so agents can track refinement, access domain data, and apply foundational scientific reasoning.Examples include AccelMat (Kumbhar et al., 2025), CellForge (Tang et al., 2025d), and BioMaster.Source: §2.2.4 · p. 20
• Current memory systems face limited context windows, information loss, parametric opacity, brittle external retrieval, and memory overload, requiring adaptive updating, lifelong learning, and efficient forgetting.These challenges concern both scalability and the maintenance of relevant information over extended research cycles.Source: M3. Intrinsic Knowledge · p. 17, §2.2.4 · pp. 20–21

2.3 Action Space

• Scientific agents’ action spaces operationalize plans through tool use, information retrieval, code execution, and cognitive reasoning, often integrating multiple action types into unified workflows.Tool use extends LLMs with real-time retrieval, precise execution, specialized computation, and experiment simulation.Source: §2.3 · p. 21, §2.3.1 · p. 21

2.3.1 T1. Tool Use / Environmental Control

• APIs and code libraries expose domain knowledge and specialized algorithms through standardized interfaces, extending scientific agents beyond training-data, expertise, and computational limitations.ToRA (Gou et al., 2024) integrates SymPy, SciPy, and CVXPY, while ChemCrow deploys 18 expert-designed tools for chemistry tasks.Source: §2.3.1 · p. 21

T3. Code Generation & Execution

• Simulation tool sets translate natural-language instructions into executable codes or control signals, enabling hypothesis testing, parameter optimization, and validation across physics, chemistry, engineering, and astronomy.Examples include SGA’s simulator-based constitutive-law discovery, MyCrunchGPT’s surrogate-assisted airfoil optimization, OpenFOAM workflows, molecular simulations, robotic synthesis, and telescope control.Source: T3. Code Generation & Execution · p. 23
• Simulation and domain-model integrations improve physical validity, computational precision, and access to specialized expertise, but face high costs, heterogeneous interfaces, limited accessibility, and difficult error handling.The survey motivates standardized interfaces, robust validation and error detection, and hybrid workflows combining fast approximate tools with precise validation.Source: T3. Code Generation & Execution · pp. 23–25, §2.3.3 · p. 28
• Domain models function as specialized tools that provide expert predictions or analyses while the agent orchestrates workflows, decisions, and result synthesis.Examples span protein design and structure prediction, material-property estimation, surrogate modeling, and AlphaFold-based validation of proposed protein sequences.Source: T3. Code Generation & Execution · p. 24

2.3.2 T2. Search & Information Retrieval

• Search and retrieval actions acquire literature, database records, and web information through query, retrieval, extraction, ranking, and synthesis, grounding agents’ reasoning in external knowledge.Effectiveness depends on query formulation and relevance assessment, while information overload, source quality, contradictory evidence, and integration complexity remain challenges.Source: §2.3.2 · pp. 25–26

2.3.3 T3. Code Generation & Execution

• Code generation and execution follow specification-synthesis-execution cycles in which agents produce programs or workflows, run them, inspect outputs and errors, and iteratively debug or refine them.Applications include machine-learning experiments, biomedical analyses, chemistry protocols, neuroscience scripts, CFD configurations, physics models, and instrument control.Source: §2.3.3 · pp. 26–28

2.3.4 T4. LLM-Based Reasoning / Cognitive Actions

• LLM-based reasoning supports hypothesis generation, creative synthesis, analytical decomposition, critical evaluation, experimental interpretation, and multi-step inference across scientific domains.These cognitive actions exploit the model’s internal capabilities rather than invoking external systems, forming the “thinking” component of scientific agents.Source: §2.3.4 · p. 28
• Prompt design activates and channels reasoning through hypothesis-generation, analytical, creative-synthesis, critical-evaluation, and chain-of-thought prompts, while capability and reliability depend on the underlying LLM and prompt quality.Representative systems use iterative idea refinement, novelty-optimized proposal generation, process-supervised reasoning, multi-agent specialization, and domain-specific reasoning chains.Source: §2.3.4 · p. 29
• Reasoning remains vulnerable to plausible but flawed outputs, absent formal correctness guarantees, and hallucinations, necessitating tool-based validation, retrieval grounding, and careful prompt engineering.These limitations constrain the reliability of flexible cognitive capabilities and natural-language scientific collaboration.Source: §2.3.4 · pp. 29–30

2.3.5 Discussion

• The four action types—tool use, search and retrieval, code generation, and LLM reasoning—address complementary limitations, and their integration can support workflows from hypothesis generation through execution and analysis.Tools decouple strategic reasoning from low-level execution by encapsulating algorithms, domain knowledge, computational resources, and, through simulators, virtual experimental validation.Source: §2.3.5 · p. 30
• Current action-space systems are constrained by static tools, changing APIs, costly or delayed resources, cascading heterogeneous-tool failures, security risks, and version-dependent reproducibility problems.Proposed remedies include runtime tool discovery and composition, middleware abstractions, standardized interfaces and error semantics, probabilistic action selection, resource-aware planning, composable primitives, and formal safety and validation protocols.Source: §2.3.5 · p. 30

2.4 Verifier

• Scientific-agent verification spans four complementary mechanisms—self-correction, multi-agent critique, human-in-the-loop oversight, and tool-based validation—often combined in layered architectures.These mechanisms assess validity, feasibility, consistency, and correctness before outputs guide experiments, resource allocation, or publication.Source: §2.4 · p. 31

2.4.1 V1. Self-Correction / Reflective Verification

• Self-correction uses iterative reflection–revision cycles to detect and repair flaws, but single-model blind spots, missing external grounding, and hallucinated critiques limit reliability.Implementations include hypothesis refinement, code debugging from execution errors, retries, self-consistency checks, and feedback-driven protocol or design revision.Source: §2.4.1 · pp. 31, 33, 34

2.4.2 V2. Multi-Agent Critique / Role-Based Verification

• Multi-agent critique assigns differentiated roles to multiple LLMs, aggregating diverse reviews for broader error coverage while incurring coordination costs, computational overhead, conflicts, and shared biases.Workflows may use adversarial debate, consensus-building, hierarchical review, specialized critics, or graph-structured message passing.Source: §2.4.2 · pp. 34–35

2.4.3 V3. Human-in-the-Loop / Expert Oversight

• Human-in-the-loop verification places experts at critical decision points for authoritative approval, rejection, feedback, safety oversight, and ambiguity resolution, but creates bottlenecks, subjectivity, and cost.Interaction ranges from continuous oversight to selective or exception-based intervention, including approval gates, collaborative iteration, and manual debugging.Source: §2.4.3 · pp. 35–37

2.4.4 V4. Tool-Based Validation / Computational Verification

• Tool-based validation grounds verification in execution, simulation, databases, and APIs, providing objective scalable signals, while tool scarcity, simulator fidelity, and computational cost remain limitations.Examples include code testing, physics or chemistry simulation, molecular validity checks, and factual cross-checking against authoritative repositories.Source: §2.4.4 · pp. 37–38

• Layered verification combines self-correction, multi-agent critique, human oversight, and tool-based validation for defense-in-depth quality assurance.Representative systems include AIGS (Liu et al., 2024c), Agent Laboratory (Schmidgall et al., 2025), AtomAgents (Ghafarollahi and Buehler, 2024a), and Coscientist.Source: §2.5 · p. 38
• Verification strategies vary with task criticality, domain maturity, cost constraints, and automation requirements, balancing human review, automated tools, lightweight correction, and expert collaboration.High-risk tasks favor human oversight, established domains enable tool-based validation, rapid iteration uses lightweight self-correction, and autonomous systems emphasize self-correction and tools.Source: §2.5 · p. 38
• Future priorities include metaverification, uncertainty-adaptive verification, verification-aware learning, and formal frameworks that provide safety guarantees for high-stakes autonomous workflows.Source: §2.5 · pp. 38–39

3 Benchmarks

• Benchmarks evaluate LLM-based scientific agents across general reasoning and domain-specific research capabilities, from basic problem-solving to paper reading and experiment design.This taxonomy is summarized in Figure 11 and structures the section’s evaluation framework.Source: §3 · p. 39, Figure 11 · p. 40

3.1 General Reasoning Ability Evaluation

• General reasoning benchmarks progress from K-12 mathematics and science to higher-education, graduate, and expert-level assessments of reasoning, knowledge retrieval, and multimodal understanding.Examples include Geometry3K, MathVista, SciBench, MME-SCI, GPQA, SuperGPQA, HLE, and other expert-driven evaluations.Source: §3.1 · pp. 39, 41

3.2 Scientific Research-Oriented Ability Evaluation

• Scientific research-oriented benchmarks assess figure comprehension, hypothesis discovery, cross-paper reasoning, end-to-end research, experimental design, workflow automation, reproducibility, security, and scientific trustworthiness.They simulate research workflows using textual, visual, empirical, and tool-mediated tasks across multiple scientific domains.Source: §3.2 · pp. 41–43

3.3 Discussion

• Together, the benchmarks cover scientific skills across research stages, from foundational reasoning through advanced hypothesis generation and experimental automation.This breadth supports comprehensive assessment and guides the development of scientific AI systems.Source: §3.3 · p. 43
• Current evaluations rely heavily on static, predefined tasks, obscure step-level failures, and struggle to standardize metrics across diverse scientific domains.Future benchmarks should be adaptive and continuously updated, support multi-turn refinement from experimental feedback, and include domain-specific and cross-disciplinary evaluation.Source: §3.3 · p. 43

4 Applications

• Scientific agents automate multiple research stages, increasing workflow efficiency and reducing resource barriers across scientific disciplines.These stages include hypothesis formulation, experimental design, planning, data analysis, and evaluation.Source: §4 · p. 43

4.1.1 Chemical Synthesis and Reaction Optimization

• LLM-based agents automate chemical synthesis planning, reaction-condition optimization, and experimental execution with minimal human intervention.Coscientist integrates LLM planning with robotic equipment, while ChemCrow invokes 18+ specialized cheminformatics tools for synthesis, drug-design, and materials tasks.Source: §4.1.1 · p. 43
• Multi-agent chemical systems coordinate literature search, protocol design, computation, robotics, reaction optimization, purification, and results analysis while supporting safety and reproducibility.ChemAgents coordinates specialized roles, and LLM-RDF automates end-to-end synthesis development from literature search through product purification.Source: §4.1.1 · pp. 44–45

4.1.2 Molecular Design and Optimization

• Molecular-design agents perform inverse design, reaction prediction, structure elucidation, quantum-chemistry workflows, and de novo generation while optimizing activity, synthesizability, and toxicity.MolRL-MGPT uses supervised and reinforcement learning with experience replay, while IR-Agent coordinates specialized agents to interpret IR spectra.Source: §4.1.2 · p. 45
• These molecular agents navigate vast chemical spaces using learned chemical principles, computational predictions, tool learning, memory, self-repair, and sufficiency validation.CheMatAgent separates tool selection from execution, and El Agente uses procedural, semantic, and episodic memory for quantum-chemistry workflows.Source: §4.1.2 · p. 45

4.1.3 Materials Discovery and Characterization

• Materials-science agents generate hypotheses, predict properties, run physics-based simulations, and validate designs through closed-loop workflows involving human oversight and physical synthesis.MAPPS unifies planning, simulations, and scientist feedback, while AtomAgents combines multimodal data integration with LAMMPS simulations.Source: §4.1.3 · pp. 45–46
• Materials agents also automate microscopy, integrate knowledge bases and computational tools, correlate multimodal data, reason over ontologies, and generate metal-organic frameworks from textual inputs.These systems span automated AFM experiments, materials question answering, structured-knowledge hypothesis generation, and MOF prediction and generation.Source: §4.1.3 · pp. 45–46

4.2.1 Drug Discovery and Biomedical Research

• Drug-discovery agents automate workflows from target identification and interaction prediction through candidate optimization, validation, and hypothesis generation using knowledge graphs, literature search, specialized roles, and expert oversight.Examples include DrugAgent, Robin, BioScientist Agent, OriGene, DrugPilot, DrugAssist, AI co-scientist, STELLA, MedAgents, and GatorTronGPT.Source: §4.2.1 · p. 46

4.2.2 Genomics and Bioinformatics Workflows

• Genomics agents automate bioinformatics workflows from planning and execution through quality control, single-cell analysis, regulatory reconstruction, and biological interpretation using retrieval, memory, debugging, code generation, and specialized tools.Biomni integrates 150+ tools, 105 software packages, and 59 databases, while GeneGPT achieves state-of-the-art genomics question-answering performance using NCBI Web APIs and augmented decoding.Source: §4.2.2 · pp. 46–47

4.2.3 Protein Engineering and CRISPR Applications

• Protein-engineering and CRISPR agents support de novo protein design, physical and machine-learning analysis, gene-editing experiment planning, molecular simulation, and virtual-cell modeling through iterative validation and specialized collaboration.Sparks discovers protein design principles through computational and experimental cycles, while CRISPR-GPT integrates expert knowledge and computational tools for guide design, delivery recommendations, protocols, and validation experiments.Source: §4.2.3 · p. 47

4.3.1 Computational Fluid Dynamics and Mechanics

• LLM agents automate OpenFOAM CFD and mechanics workflows, reducing expertise barriers by generating configurations, executing simulations, interpreting results, and coordinating iterative multi-agent refinement.Systems use retrieval over OpenFOAM documentation, specialized knowledge indices, reviewer agents, self-correction, division of labor, and reinforcement learning for autonomous fluid-dynamics exploration.Source: §4.3.1 · pp. 47–48

4.3.2 Electromagnetic Fields and Quantum Systems

• Specialized agents bridge high-level scientific goals and low-level implementation in electromagnetic and quantum systems, including EM-wave control, metamaterial optimization, and autonomous superconducting-processor operation.k-agents demonstrated autonomous calibration and operation of a superconducting quantum processor at human-level performance, while MetaAgent coordinated specialized EM agents and SGA paired LLM ideas with physics simulators.Source: §4.3.2 · p. 48

4.4.1 Astronomical Observation and Telescope Control

• Observation agents translate astronomers’ natural-language objectives into verified telescope-control sequences, execute approved programs, adapt to unexpected situations, and maintain safety constraints for expensive instruments.StarWhisper generates pointing coordinates, exposure times, and filter selections, while LLMSat supports autonomous space-exploration scenarios and Mephisto performs iterative multi-band galaxy SED modeling.Source: §4.4.1 · p. 48

4.4.2 Cosmological Data Analysis and Modeling

• Cosmological analysis agents use specialized multi-agent systems, astronomy-trained language models, mathematical models, and end-to-end simulation workflows to analyze datasets, answer astronomical questions, and explore celestial dynamics.AstroSage-Llama-3.1-70B underwent continued pretraining using 168,000 GPU-hours. CosmoAgent models civilization interactions, while ASA autonomously designs, executes, and reports on celestial-mechanics simulations.Source: §4.4.2 · p. 49
• LLM agents automate data-intensive astronomical research from hypothesis formulation and computational analysis to scientific interpretation, enabling astronomers to explore larger datasets and test more hypotheses than manual analysis allows.These systems support autonomous exploration of cosmological datasets, simulation design and execution, and research reporting.Source: §4.4.2 · p. 49

4.5.1 Geospatial Analysis, and Geological/Climate Modeling

• Geospatial agents automate spatial analysis and data retrieval by combining code interpreters, static analysis, retrieval, Monte Carlo Tree Search, error tracing, and self-debugging modules, including GeoAgent (Chen et al., 2024a) and LLM-Find (Ning et al., 2025).These systems iteratively refine generated code and retrieve from scalable sources such as OpenStreetMap and US Census data.Source: §4.5.1 · pp. 49–50
• Geological and mineral-exploration agents support map interpretation, remote-sensing assessment, and domain-specific reasoning through expert-agent consultation, hierarchical judging, multimodal integration, proprietary data, and knowledge graphs.GeoMap-Agent (Huang et al., 2025b), MineAgent (Yu et al., 2024), and GeoMinLM exemplify these approaches.Source: §4.5.1 · pp. 49–50
• Simulation-oriented systems translate natural-language requests into technically constrained geomechanical workflows, while PANGAEA GPT (Pantiukhin et al., 2025) uses multi-agent interaction to improve geoscientific data processing and collaboration.GeoSim.AI uses retrieval-augmented geomechanics knowledge and data/tool bases to maintain physical plausibility of parameters and configurations.Source: §4.5.1 · p. 50
• Earth-Agent (Feng et al., 2025b) enables cross-modal, multi-step, quantitative spatiotemporal reasoning over RGB and spectral data, while ClimateGPT (Thulke et al., 2024) supports climate question answering and literature and dataset retrieval.Together, these systems extend agentic capabilities across Earth observation and climate science applications.Source: §4.5.1 · p. 50

4.6.1 Automated Machine Learning Research

• Automated ML research agents support the full research pipeline from idea generation and experimentation through analysis, paper writing, peer review, and iterative feedback, but remain limited mainly to well-defined computational domains.AI Scientist-v2 adds agentic tree search and VLM-based figure refinement, while Agent Laboratory, MLR-Copilot, AIGS, and CycleResearcher incorporate multi-agent exploration, feedback, falsification, or preference training.Source: §4.6.1 · p. 50

4.6.2 Algorithm Discovery and Optimization

• Algorithm-focused agents use evolutionary search, specialized tools, iterative critique, and automated execution feedback to discover algorithms, optimize implementations, and generate novel solutions across diverse scientific and engineering domains.AlphaEvolve evolves code for scheduling, matrix multiplication, circuit design, and compiler improvement, while SciAgent, HyperGen, and VirSci extend tool-augmented reasoning and iterative scientific idea generation.Source: §4.6.2 · pp. 50–51

4.6.3 Mathematical Reasoning and Formal Methods

• Mathematical reasoning agents improve multi-step problem solving through tool integration, imitation learning, supervised and reinforcement learning, preference optimization, and context-pruning strategies.Tool integration and verification partially address correctness limitations by grounding reasoning in symbolic computation, though reliable guarantees and complex multi-step proofs remain unresolved.Source: §4.6.3 · p. 51

4.7.1 Literature Analysis and Research Idea Generation

• Literature and idea-generation agents automate scholarly search, reference selection, novelty assessment, hypothesis refinement, and research coordination, augmenting scientific creativity while retaining the need for human evaluation.Examples include reinforcement-learning search, literature-grounded ideation, Chain-of-Ideas reasoning, academic-graph integration, closed-loop research workflows, and uncertainty-reduction frameworks.Source: §4.7.1 · p. 52

4.7.2 Specialized Domain Applications and Platforms

• Specialized agents apply conversational interfaces, retrieval, code execution, and robotic automation to food science, neuroscience, SciML, bioinformatics, and chemistry.These systems translate natural-language requests into analyses or experiments, including flavor discovery, animal-behavior analysis, airfoil and fluid simulations, laboratory automation, and parallel chemistry workflows.Source: §4.7.2 · p. 52

Applications

• LLM-based scientific agents can enhance data interpretation, support complex decision-making, and generate novel hypotheses across fields including biomedical analysis and materials science.These capabilities may accelerate scientific discovery.Source: §4.8 · pp. 52–53

Limitations

• Current applications are often domain-specific, rely on static models that cannot adapt to real-time data or evolving challenges, and lack sufficient validation against scientific benchmarks.These limitations raise concerns about generalizability, reproducibility, and reliability.Source: §4.8 · p. 53

Future Directions

• Future systems should integrate heterogeneous data, support cross-disciplinary collaboration, detect errors in real time, use adaptive feedback, and employ multimodal LLM architectures.These enhancements are intended to improve system robustness.Source: §4.8 · p. 53
• Collaboration between domain experts and AI researchers is crucial for fine-tuning scientific agents’ decision-making so outputs align with established scientific principles and practices.This collaboration addresses the need for more reliable scientific-agent outputs.Source: §4.8 · p. 53

5 Ethics

• Unchecked AI agency can endanger public safety and scientific integrity, motivating non-agentic Scientist AI designs that prioritize explanation and transparency, alongside principled oversight, provenance-rich workflows, and transparent documentation.Tang et al. (2025c) argue that autonomy itself introduces systemic risks and that safeguarding should take priority over capability expansion.Source: §5 · p. 53

5.1 Agency and Autonomy

• Scientific agents should remain bounded tools under human supervision, using ethical rules, feedback, refusal policies, and safety evaluation to prevent deceptive or persistent goals.SafeScientist combines rule-based refusals with an ethical-reviewer agent, while SciSafetyBench evaluates whether agents detect and decline unsafe or unethical tasks.Source: §5.1 · p. 53

5.2 Transparency and Explainability

• Transparent scientific agents must record complete operational traces, including queries, tool calls, code, retrieval logs, scripts, and intermediate results, enabling replay, auditing, and reproducibility.ScienceBoard records agent interactions so experiments can be replayed and verified, extending explainability beyond verbal rationales.Source: §5.2 · pp. 53–54

5.3 Hallucinations and Reliability

• Reliability requires architectural verification because hallucinations from poor data, ambiguous prompts, over-generalization, or manipulated prompts can mislead scientific reasoning and public trust.Verifiers, process supervision, iterative feedback, and memory of validated evidence turn reliability into an active self-regulating process.Source: §5.3 · p. 54

5.4 Vulnerability and Security

• Scientific agents face prompt injection, model extraction, poisoned knowledge graphs, memory poisoning, tool-response injection, data leakage, and collusion across multi-agent systems.Countermeasures include capability firewalls, context sanitization, strict access controls, encryption, and cryptographic audit logs.Source: §5.4 · p. 54

5.5 Bias, Fairness, and Data Integrity

• Fair scientific inference depends on diverse data, transparent provenance, reproducibility, bias audits, fairness-aware algorithms, and retrieval strategies that seek contradictory findings.BioAgents and ChemDFM couple retrieval logs with validation layers to document evidence flow.Source: §5.5 · p. 54

5.6 Accountability and Governance

• Accountability should combine technical verifiers, human supervisors, and institutional policies, while final moral responsibility remains with humans rather than autonomous agents.Multi-Agent Ethics Advocate, SafeScientist, and RAS-Eval embed reviewer roles, refusals, and standardized safety testing into governance.Source: §5.6 · p. 54

5.7 Intellectual Property and Research Integrity

• Ethical AI-assisted research requires provenance metadata for datasets, code, and text, including model versions, prompts, human contributors, and licensing terms.Immutable records, digital signatures, disclosure policies, and standardized reporting protect intellectual property and reproducibility.Source: §5.7 · p. 55

6 Conclusion

• Scientific agents achieve effectiveness through synergistic integration of planners, memory, action spaces, and verification rather than isolated component sophistication.Planners decompose tasks, memory supports iterative hypothesis refinement, action spaces connect tools and reasoning, and verification enables self-correction, critique, oversight, and validation.Source: §6 · p. 55
• Benchmarks must assess both general reasoning and domain-specific scientific competence, while applications and ethics determine responsible real-world impact.The survey emphasizes innovation across disciplines alongside reproducibility, transparency, and adherence to stringent research standards.Source: §6 · p. 55
• The survey projects that continued advances in scientific agents could revolutionize discovery by strengthening interdisciplinary collaboration and innovation.This future depends on bridging theoretical research with practical applications.Source: §6 · p. 55

A Classification of All Related Work

• The taxonomy classifies scientific agents by planner, memory, action, and verifier mechanisms across chemistry, materials, life, biomedical, physical, environmental, computational, and meta-research domains.Its categories include instructional, context-augmented, deliberative, search-based, multi-agent, programmatic, and learned planners; historical, external, and parametric memory; tool, retrieval, code, and reasoning actions; and self-correction, critique, human oversight, and tool validation.Source: Table 4 · p. 56, Table 5 · p. 57, Table 6 · p. 58