Agent4S: The Transformation of Research Paradigms from the Perspective of Large Language Models

Boyuan Zheng; Cunshi Wang; Fangfu Ye; Jiaji Wu; Jia Li; Jianwei Shuai; Lei Lei; Mengwei Qu; Mingzhou Tan; Rui Wang

REVIEW 3 major objections 5 minor 11 references

Agent4S: The Transformation of Research Paradigms from the Perspective of Large Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper argues that the era of AI as a mere analysis tool is ending, and that LLM-driven agents that run the entire research workflow—from planning to experiment to interpretation—are the true Fifth Scientific Paradigm, organized into…

desk verdict A useful five-level agent taxonomy for AI4S, but the 'true Fifth Paradigm' claim is unsupported and the paper contradicts its own definition of a paradigm shift. read the letter →

arxiv 2506.23692 v1 pith:MEI7EXTG submitted 2025-06-30 cs.AI

Boyuan Zheng , Zerui Fang , Zhe Xu , Rui Wang , Yiwen Chen , Cunshi Wang , Mengwei Qu , Lei Lei

show 8 more authors

Zhen Feng Yan Liu Yuyang Li Mingzhou Tan Jiaji Wu Jianwei Shuai Jia Li Fangfu Ye

This is my paper · ORCID

classification cs.AI

keywords AgentforScienceAI4SFifthScientificParadigmLLM-drivenagentsresearchautomationAIScientistmulti-agentcollaborationdataprocessing

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the next scientific paradigm will come not from better AI analysis tools but from LLM-driven agents that run the research process itself. It calls this vision Agent for Science (Agent4S) and positions it as the true Fifth Scientific Paradigm, contrasting it with AI4S, where AI remains a data-analysis method inside the fourth, data-driven paradigm. To make the vision concrete, the paper defines five levels of Agent4S, from automating a single tool (L1) to fully autonomous single-laboratory research (L4) and cross-laboratory collaboration of AI Scientists (L5). A sympathetic reader would care because the five-level hierarchy gives researchers a route map and names the technical bottlenecks, such as memory, planning, tool-invocation protocols, embodied intelligence, and agent-to-agent communication, that must be solved at each step.

What carries the argument

The central object is the five-level Agent4S hierarchy, a classification that ties the evolution of agent technology to the degree of research automation. The hierarchy carries the argument by specifying, at each level, the underlying agent technology, the affected research phase, and the implementation challenge: prompt engineering plus function calling at L1, workflow orchestration at L2, reasoning with context engineering and the Model Context Protocol (MCP) at L3, embodied and hardware-integrated intelligence at L4, and agent-to-agent (A2A) protocols at L5. This structure separates what the paper calls mere improvements within the data-driven paradigm from a genuine paradigm shift.

What would settle it

A concrete test: run an L3-level agent in an automated laboratory on a bounded research question and count how many complete hypothesis-experiment-analysis loops it closes without human replanning. If it closes none, the leap from L3 to L4 is not just a scaling problem; if it closes even one full loop, the roadmap receives direct support.

Watch

Extended reading notes

Core claim

The paper's central claim is that the defining feature of a scientific paradigm is how data are acquired and processed, and that agents change both at once. Under this view, AI4S belongs inside the fourth, data-driven paradigm because its algorithms only process data, while Agent4S replaces the human-driven loop with agents that plan, invoke tools, execute experiments, and interpret results. The paper classifies Agent4S into L1 (automation of a single scientific tool), L2 (automation of complex scientific pipelines), L3 (intelligent single-flow research with a reasoning agent), L4 (full-process intelligence within one laboratory), and L5 (multiple intelligent processes collaborating across laboratories via agent-to-agent protocols). The authors assert that L4 and L5, not today's tools, are the final form of the fifth paradigm.

Load-bearing premise

The roadmap assumes that making agents bigger and better connected, through longer memory, longer planning, better tool use, physical embodiment, and inter-agent communication, is enough to let machines run whole research projects on their own, something no one has shown yet.

Editorial extensions

If this is right

If Agent4S is the Fifth Scientific Paradigm, the unit of progress in science shifts from analysis algorithms to agents that own the research loop, and AI4S becomes a component technology inside it.
At L3, domain scientists would work alongside AI Scientists that autonomously plan, invoke tools, analyze data, and iterate within a single workflow.
At L4, an agent would act as the coordinator of a full research project inside one laboratory, covering question formulation, design, hypothesis, experiment, and interpretation.
At L5, agents across laboratories would form an interdisciplinary network using agent-to-agent protocols, making cross-laboratory collaboration the default organizational form.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the paper leaves implicit is that research workflows could be graded by the level at which a human must re-enter the loop, allowing laboratories to measure their current position on the L1–L5 scale.
The paper's framing implies, without saying so, that the binding constraints on the fifth paradigm are not model intelligence alone but agent memory, tool protocols, laboratory hardware integration, and cross-agent communication standards.
The taxonomy could be sharpened by mapping it to capability tiers similar to autonomous-driving levels, where each tier has explicit take-over conditions; that mapping is an extension, not part of the paper.
If L4 and L5 are achieved, the unit of scientific credit may shift from the human author to the human-agent team, since the agent would initiate as well as execute research; the paper does not discuss this.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the Agent4S taxonomy (L1–L5, from single-tool automation to multi-laboratory agent networks) is a genuinely useful way to talk about where agent technologies sit in scientific research, and the mapping to concrete technologies (function calling, workflows, MCP, A2A) is helpful. If that were the whole paper, I'd be moderately positive. But the paper wraps it in a bigger claim—that Agent4S is 'the true Fifth Scientific Paradigm'—and that claim does not survive contact with the paper's own definition.

Section 1 says paradigm shifts are 'essentially transformations in methods of data acquisition and data processing.' Section 3.1 then says L1 and L2 'involve the automation of fixed processes' and 'remain tools within the research workflow.' Automating a fixed process changes who executes the steps, not the method. So under the paper's own criterion, L1/L2 are productivity improvements inside the fourth paradigm, not components of a fifth one. That leaves the taxonomy as an automation-maturity scale for tools, which is fine, but not a new scientific paradigm. The abstract overclaims.

What's actually new: the five-level classification itself, and the way it links each level to the agent technology that enables it. I don't know of another paper that lays out that mapping in quite this way. The authors also correctly distinguish AI4S as an analytical method from Agent4S as an orchestration layer, and Section 4's contrast between the two is the clearest part of the paper.

Soft spots, in order of importance: (1) the internal criterion mismatch above; (2) the roadmap from L3 to L4/L5 rests on an unvalidated scaling assumption—there is no evidence that longer memory, better planning, and MCP invocation will be sufficient for full lab autonomy; (3) the paper leaves explicit 'specific cases and references needed' notes in the text, which is honest but means the illustrative examples are not yet filled in; (4) it cites no prior automation-maturity taxonomies, even though Section 4 acknowledges they exist.

I don't think the central argument holds as written. For a serious referee, this is a paper that could be revised into something valuable, but the 'true Fifth Paradigm' claim needs to be moderated or the definitional inconsistency resolved. As it stands, I'd treat it as a proposal for a taxonomy, not a demonstration. My recommendation: send it to peer review, but with a strong expectation of major revision. The taxonomy deserves discussion; the paradigm claim does not.

Referee Report

3 major / 5 minor

Summary. The paper proposes "Agent for Science" (Agent4S), defined as LLM-driven agents automating the entire research workflow, and claims it is "the true Fifth Scientific Paradigm." It introduces a five-level hierarchy (L1–L5) from single-tool automation to multi-laboratory multi-agent collaboration, maps each level to current agent technologies (function calling, workflow orchestration, MCP, A2A), and distinguishes Agent4S from AI4S by casting AI4S as a component of the data-analysis stage. The paper contains no experiments, derivations, or quantitative evaluation; it is a conceptual position piece that presents a taxonomy and a roadmap.

Significance. The five-level hierarchy is clear and potentially useful as an organizing framework for discussing increasing automation and intelligence in scientific research. The explicit mapping to concrete technologies such as MCP and A2A gives practitioners a vocabulary for maturity levels, and the distinction between AI4S-as-data-analysis-method and Agent4S-as-productivity-tool is a useful clarification. However, the central claim that Agent4S is "the true Fifth Scientific Paradigm" is asserted rather than demonstrated: the paper's own criterion for a paradigm shift, stated in Section 1, is not consistently applied to its own levels, and the roadmap from L3 to L5 rests on extrapolation from early-stage technology. As presented, the hierarchy is better supported as an automation-maturity scale or a research agenda than as the definitive characterization of a new scientific paradigm.

major comments (3)

[Section 1 / Section 3.1 / Table 2] The paper's working definition, given in Section 1, is that paradigm revolutions are "essentially transformations in methods of data acquisition and data processing." In Section 3.1 and Table 2, L1 and L2 are described as "automation of fixed processes" that "remain tools within the research workflow." Automating a fixed process does not change the method by which data are acquired or processed; it changes who or what executes the steps. Under the Section 1 criterion, L1 and L2 are productivity improvements within the fourth paradigm, not components of a fifth. If L1 and L2 are excluded, the five-level hierarchy becomes an automation-maturity scale rather than a scientific paradigm; if they are included, the paper abandons its own criterion without saying so. The abstract's categorical claim that Agent4S is "the true Fifth Scientific Paradigm" is therefore not supported by the paper's own text.
[Section 3.1] The manuscript itself flags missing empirical grounding: after describing L1 and L2, the text reads "(specific cases and references needed, note: the implementation must be a single agent)" and "(specific cases and references needed, note: implementation must be workflow-driven multi-agent ...)". These are not presentation issues. The L1/L2 definitions are the base of the five-level roadmap, and without concrete, cited examples the hierarchy cannot be checked against existing systems and the boundary between L1 and L2 remains unverifiable. The authors should either supply real instances with references or explicitly recast the paper as a proposal whose levels await operationalization.
[Section 3.1 (L3–L5)] The transition from L3 to L4 and L5 rests entirely on extrapolation: "Looking ahead, as Agent capabilities in memory length, multi-step planning, and MCP invocation continue to advance" and "with the advancement of full-process intelligence in each laboratory." No evidence or detailed mechanism is given to show that these capabilities are sufficient for full research-process autonomy, including hypothesis generation, experiment design, and cross-laboratory collaboration. This assumption is legitimate for a roadmap, but it cannot ground the paper's claim that Agent4S currently constitutes the Fifth Paradigm. The authors should either present supporting evidence or qualify the claim as a forward-looking vision rather than an established paradigm.

minor comments (5)

[Abstract] The informal contraction "doesn't" should be replaced with "does not" for a formal research paper.
[References] Reference [7] appears to be misattributed: the cited title "Adaptive control processes: a guided tour" is by R. Bellman, and the author entry should be corrected.
[Table 2] The "Implementation Challenges" entries for L1 and L2 ("Hardware Digitization" and "Data Transmission Robustness") are not defined or elaborated in the text; please add brief explanations of what these challenges mean.
[Figures 1 and 2] Figures 1 and 2 are referenced but the prose does not describe their contents in sufficient detail; the captions should state what each panel depicts, especially the role assignments in Figure 2.
[Section 4] The claim that previous taxonomies were "more as automation ratings rather than providing actionable technical roadmaps" is not accompanied by any specific citation or comparison; please cite and discuss at least one prior classification to substantiate the novelty claim.

Circularity Check

2 steps flagged · score 4.0 of 10

Partial definitional circularity: the 'Fifth Paradigm' conclusion is encoded in the definition of Agent4S, while L1/L2 are simultaneously labeled as paradigm transformation and as fixed tools inside the workflow.

self definitional [Abstract; Section 1; Section 4]
"The four scientific paradigm revolutions ... are essentially transformations in methods of data acquisition and data processing. ... Agent4S signifies agent-driven automation and intelligence in scientific research, emerging as a transformative productivity paradigm for data acquisition and processing."

The premise used to identify paradigm shifts is the same phrase used to define Agent4S. Section 1 says a paradigm revolution is a transformation in data acquisition/processing; Section 4 says Agent4S is 'a transformative productivity paradigm for data acquisition and processing.' Thus the conclusion 'Agent4S is the Fifth Scientific Paradigm' is already packed into the definition, and the paper offers no independent evidence that LLM-agent automation changes the method of data acquisition/processing rather than merely automating existing steps.
other [Section 3.1 and Table 2]
"Levels L1 and L2 involve the automation of fixed processes, primarily applied to the historically machine-driven data generation/acquisition stages. Fundamentally, they remain tools within the research workflow."

Table 2 lists L1 and L2 under 'Paradigm Transformation Mode' as 'Automated Tools for Fixed Processes,' while the text says they 'remain tools within the research workflow.' Under the paper's own Section 1 criterion, a paradigm transformation must change the method of data acquisition/processing; automating a fixed process does not. Counting L1/L2 as part of the fifth paradigm is therefore a labeling decision rather than a consequence of the paper's stated criterion, and it lets the five-level hierarchy appear to cover a paradigm shift when part of it is explicitly tool-level automation.

full rationale

The paper contains no empirical predictions, no fitted parameters, and no load-bearing self-citation chain; reference [11] is an independent prior proposal of the Fifth Paradigm. The central circularity is definitional: the paper's criterion for a paradigm revolution (a transformation in data acquisition/processing methods) is restated as the defining property of Agent4S, so the headline 'true Fifth Scientific Paradigm' claim is partly true by construction. The internal L1/L2 discussion reinforces this by labeling fixed-process automation as a 'Paradigm Transformation Mode' while simultaneously admitting those levels 'remain tools within the research workflow.' However, the five-level taxonomy itself, the mapping of levels to MCP, A2A, reasoning frameworks, and the distinction from classical AI4S are checkable external content, so the circularity is partial rather than total. Score 4 reflects a central claim that is partly definitional but supported by a framework with independent, inspectable substance.

Assumptions & free parameters 0 free parameters · 4 assumptions · 2 invented entities

The paper introduces no fitted parameters. It relies on several domain assumptions about what constitutes a paradigm and about the scalability of agent technologies. The invented entities are conceptual constructs rather than physical ones, and neither has independent empirical support.

assumptions (4)

domain assumption The four historical scientific paradigm revolutions are essentially transformations in data acquisition and processing methods.
Section 1: this framing is the basis for claiming Agent4S qualifies as a new paradigm because it changes data processing to agent-driven automation.
domain assumption The deeper contradiction in current research is between the growing volume of scientific information and the inefficiency of existing research paradigms.
Section 1: this motivates the need for a new paradigm; it is asserted, not measured.
ad hoc to paper The Fifth Scientific Paradigm concept proposed in reference [11] is valid but lacked a strict definition; Agent4S provides that strict definition.
Section 4: the paper adopts [11]'s characteristics as ground truth and claims its framework finally defines the paradigm.
ad hoc to paper Current and near-term agent technologies (ReAct, Tree-of-Thought, Graph-of-Thought, MCP, A2A) will scale to full research-process autonomy (L4, L5).
Section 3.1: the roadmap from L3 to L5 is an extrapolation, not a demonstrated capability.

invented entities (2)

Agent4S (agent-driven scientific research)
purpose: To name the proposed fifth paradigm and distinguish agent workflow automation from AI4S analytics.
Defined by the authors; no falsifiable prediction outside the paper.
AI Scientist (L3/L4 super-agent)
purpose: Projected future agent with autonomous planning and long-term memory that plans, runs, and interprets experiments.
Currently a forecast, not demonstrated; no external benchmark provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Agent4S: The Transformation of Research Paradigms from the Perspective of Large Language Models." pith.science (2026). https://pith.science/paper/MEI7EXTG

@misc{pith2026250623692,
  author       = {Pith},
  title        = {Pith review of: Agent4S: The Transformation of Research Paradigms from the Perspective of Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MEI7EXTG}},
  note         = {Machine review of arXiv:2506.23692}
}

read the original abstract

While AI for Science (AI4S) serves as an analytical tool in the current research paradigm, it doesn't solve its core inefficiency. We propose "Agent for Science" (Agent4S)-the use of LLM-driven agents to automate the entire research workflow-as the true Fifth Scientific Paradigm. This paper introduces a five-level classification for Agent4S, outlining a clear roadmap from simple task automation to fully autonomous, collaborative "AI Scientists." This framework defines the next revolutionary step in scientific discovery.

Figures

Figures reproduced from arXiv: 2506.23692 by the authors.

**Figure 1.** Technical framework task flow; or chaining structure generation, first-principles calculations, and database storage into one-click scripts for high-throughput materials computation (specific cases and references needed, note: implementation must be workflow-driven multi-agent, where workflows can be linear, parallel, or autonomously planned). We define this as Level 2 (L2) of Agent4S—Automation of Complex Scientif… view at source ↗

**Figure 2.** Five levels of Agent4S to an AI project leader/laboratory director that oversees the entire research lifecycle, from proposal development to completion, with full situational awareness and autonomous decision-making capabilities. L5 enables cross-disciplinary collaboration across multiple laboratories, where information flows through a network of super-agents, facilitating seamless interaction and knowledge exchang… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 10 canonical work pages

[1]

The Fourth Paradigm: Data-Intensive Scientific Dis- covery

Hey T, Tansley S, Tolle K, and Gray J. The Fourth Paradigm: Data-Intensive Scientific Dis- covery. Microsoft Research, 2009. url: https : / / www . microsoft . com / en - us / research / publication/fourth-paradigm-data-intensive-scientific-discovery/

work page 2009
[2]

Instauratio Magna (Novum Organum)

Bacon F. Instauratio Magna (Novum Organum). Publisher Unknown, 1620

work page
[3]

Planetary Astronomy from the Renaissance to the Rise of Astrophysics, Part A, Tycho Brahe to Newton

Taton R, Wilson C, and Hoskin M. Planetary Astronomy from the Renaissance to the Rise of Astrophysics, Part A, Tycho Brahe to Newton. General History of Astronomy. Cambridge University Press, 2003. url: https://books.google.com/books?id=hMgXh8jMSGgC

work page 2003
[4]

A letter of Mr

Newton I. A letter of Mr. Isaac Newton, Professor of the Mathematicks in the University of Cambridge; containing his new theory about light and colors: sent by the author to the publisher from Cambridge, Febr. 6. 1671/72; in order to be communicated to the R. Society. Philosophical Transactions of the Royal Society of London 1672;6:3075–87

work page
[5]

Cavendish H. XXI. Experiments to determine the density of the earth. Philosophical Transac- tions of the Royal Society of London 1798;88:469–526

work page
[6]

eScience-A transformed scientific method

Gray J and Szalay A. eScience-A transformed scientific method. Presentation to the Computer Science and Technology Board of the National Research Council 2007

work page 2007
[7]

Adaptive control processes: a guided tour (R

Hammer P. Adaptive control processes: a guided tour (R. Bellman). 1962

work page 1962
[8]

Breaking the curse of dimensionality with convex neural networks

Bach F. Breaking the curse of dimensionality with convex neural networks. Journal of Machine Learning Research 2017;18:1–53

work page 2017

Show all 11 references

[9]

Scientific discovery in the age of artificial intelligence

Wang H, Fu T, Du Y, et al. Scientific discovery in the age of artificial intelligence. Nature 2023;620:47–60

2023
[10]

Ai agents vs

Sapkota R, Roumeliotis KI, and Karkee M. Ai agents vs. agentic ai: A conceptual taxonomy, applications and challenge. arXiv preprint arXiv:2505.10468 2025

2025
[11]

AI4R: The fifth scientific research paradigm

LI G. AI4R: The fifth scientific research paradigm. Bulletin of Chinese Academy of Sciences (Chinese Version) 2024;39:1–9. 10

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

The Fourth Paradigm: Data-Intensive Scientific Dis- covery

Hey T, Tansley S, Tolle K, and Gray J. The Fourth Paradigm: Data-Intensive Scientific Dis- covery. Microsoft Research, 2009. url: https : / / www . microsoft . com / en - us / research / publication/fourth-paradigm-data-intensive-scientific-discovery/

work page 2009

[2] [2]

Instauratio Magna (Novum Organum)

Bacon F. Instauratio Magna (Novum Organum). Publisher Unknown, 1620

work page

[3] [3]

Planetary Astronomy from the Renaissance to the Rise of Astrophysics, Part A, Tycho Brahe to Newton

Taton R, Wilson C, and Hoskin M. Planetary Astronomy from the Renaissance to the Rise of Astrophysics, Part A, Tycho Brahe to Newton. General History of Astronomy. Cambridge University Press, 2003. url: https://books.google.com/books?id=hMgXh8jMSGgC

work page 2003

[4] [4]

A letter of Mr

Newton I. A letter of Mr. Isaac Newton, Professor of the Mathematicks in the University of Cambridge; containing his new theory about light and colors: sent by the author to the publisher from Cambridge, Febr. 6. 1671/72; in order to be communicated to the R. Society. Philosophical Transactions of the Royal Society of London 1672;6:3075–87

work page

[5] [5]

Cavendish H. XXI. Experiments to determine the density of the earth. Philosophical Transac- tions of the Royal Society of London 1798;88:469–526

work page

[6] [6]

eScience-A transformed scientific method

Gray J and Szalay A. eScience-A transformed scientific method. Presentation to the Computer Science and Technology Board of the National Research Council 2007

work page 2007

[7] [7]

Adaptive control processes: a guided tour (R

Hammer P. Adaptive control processes: a guided tour (R. Bellman). 1962

work page 1962

[8] [8]

Breaking the curse of dimensionality with convex neural networks

Bach F. Breaking the curse of dimensionality with convex neural networks. Journal of Machine Learning Research 2017;18:1–53

work page 2017