Source-linked AI summary
Foundation Models for Scientific Discovery: From Paradigm Enhancement to Paradigm Transition
Fan Liu, Jindong Han, Tengfei Lyu, Weijia Zhang, Zhe-Rui Yang, Lu Dai, Cancheng Liu, Hao Liu
TL;DR
Complex, open-ended scientific problems expose limits in existing methods, raising the question of whether foundation models merely improve scientific workflows or reshape science itself. The paper proposes a three-stage framework and reviews FM applications, arguing that these systems are beginning to catalyze a paradigm transition.
Problem
Complex and adaptive scientific problems, combinatorial search spaces, and rapidly accumulating data challenge existing methods and their simplifying assumptions.
Source: §1 · pp. 1–2Method
The paper proposes a three-stage framework for FM roles in science and reviews their applications across established paradigms, while identifying risks and research directions.
Source: §1 · p. 3Results
The paper argues that FMs are beginning to reshape the scientific paradigm, progressing from workflow infrastructure through human–AI co-creation toward autonomous discovery.
Source: §1 · pp. 2–3, §3 · p. 5Takeaways & Limitations
Abstract
from arXiv · showhide
Foundation models (FMs), such as GPT-4 and AlphaFold, are reshaping the landscape of scientific research. Beyond accelerating tasks such as hypothesis generation, experimental design, and result interpretation, they prompt a more fundamental question: Are FMs merely enhancing existing scientific methodologies, or are they redefining the way science is conducted? In this paper, we argue that FMs are catalyzing a transition toward a new scientific paradigm. We introduce a three-stage framework to describe this evolution: (1) Meta-Scientific Integration, where FMs enhance workflows within traditional paradigms; (2) Hybrid Human-AI Co-Creation, where FMs become active collaborators in problem formulation, reasoning, and discovery; and (3) Autonomous Scientific Discovery, where FMs operate as independent agents capable of generating new scientific knowledge with minimal human intervention. Through this lens, we review current applications and emerging capabilities of FMs across existing scientific paradigms. We further identify risks and future directions for FM-enabled scientific discovery. This position paper aims to support the scientific community in understanding the transformative role of FMs and to foster reflection on the future of scientific discovery. Our project is available at https://github.com/usail-hkust/Awesome-Foundation-Models-for-Scientific-Discovery.
1 Introduction
Scientific discovery has advanced through experimental, theoretical, computational, and data-driven paradigms, but complex, adaptive problems expose limits in current methods. This paper argues that foundation models are beginning to reshape science and proposes a three-stage framework for that evolution.
1 Introduction
- Complex, open-ended scientific problems and growing data expose a mismatch between scientific challenges and available methods.Examples include consciousness, protein-folding pathways, social polarization, and combinatorial search spaces; many computational models also rely on assumptions misaligned with dynamic, nonlinear systems.Source: §1 · pp. 1–2
- Foundation models’ adaptability across language, code, and multimodal inputs prompts the question of whether they enhance existing science or catalyze a new paradigm.The paper contrasts supporters who see changes to scientific discovery with skeptics who view FMs as tools that amplify existing methods.Source: §1 · p. 2
- The paper proposes a progression from workflow infrastructure, through human–AI co-creation, to autonomous scientific discovery.The stages move from supporting established epistemic structures to systems that may initiate questions, execute research, and generate knowledge with minimal human intervention.Source: §1 · pp. 2–3
- Its contributions are a conceptual framework, a systematic review across four scientific paradigms, and a research agenda addressing risks and future directions.The agenda includes aligning epistemic goals with emerging AI capabilities.Source: §1 · p. 3
2 Background and Preliminary
Scientific discovery has developed through experimental, theoretical, computational, and data-driven paradigms, each expanding what researchers can study while bringing distinct limits. Foundation models are general-purpose systems that transfer learned knowledge across tasks and modalities.
A Brief History of Scientific Discovery
- Experimental, theoretical, computational, and data-driven approaches successively expanded scientific discovery, but each has limitations, especially in isolation.Experimental work can struggle with inaccessible systems, theories with empirical testability, and data-driven methods with causal inference, interpretability, and robustness under distribution shift.Source: A Brief History of Scientific Discov- · pp. 3–4
Foundation Models
- Foundation models are large-scale neural networks trained on diverse data and adapted to downstream tasks through prompting or fine-tuning.Their pretraining supports transfer across domains and modalities, which is valuable where scientific labels are scarce and tasks are open-ended.Source: A Brief History of Scientific Discov- · p. 4
- Examples such as GPT-4, CLIP, AlphaFold, and FunSearch illustrate FM use in language, zero-shot image classification, protein folding, and mathematical discovery.CLIP was trained on 400 million image–text pairs and can classify images from natural-language prompts without additional fine-tuning.Source: A Brief History of Scientific Discov- · p. 4
3 Rethinking Scientific Paradigms in the Era of Foundation Models
The paper frames FM-enabled scientific discovery as a progression in autonomy and task scope, from infrastructure within established practice to collaboration and prospective autonomous discovery.
Meta-Scientific Integration
- Meta-Scientific Integration uses FMs to coordinate workflows and improve efficiency while leaving objectives and scientific reasoning under human control.These systems automate tasks such as preprocessing and literature retrieval but do not initiate or reframe inquiry.Source: §3 · pp. 4–5
Hybrid Human-AI Co-Creation
- Hybrid Human-AI Co-Creation makes FMs collaborators in question generation, hypothesis development, experiment planning, and interpretation, with humans still initiating and validating their work.FMs have moderate autonomy within scoped environments, while human oversight and scientific norms remain boundaries.Source: §3 · p. 5
Autonomous Scientific Discovery
- Autonomous Scientific Discovery envisions agents initiating questions, conducting the research cycle, interpreting outcomes, and updating models with minimal or no human oversight.The paper calls this prospective machine-led stage a fifth scientific paradigm, while noting that it remains formative and raises questions about responsibility, trust, and validation.Source: §3 · p. 5
4 Foundation Model Integration Across Scientific Paradigms
The review describes FM applications across experimental, theoretical, computational, and data-driven science, as well as workflows connecting these approaches. Across these areas, FMs support design, reasoning, modeling, inference, and coordination.
4.1 Experiment-Driven Paradigm
- In experimental science, FMs support resource-aware experiment design and execution, from guiding Bayesian optimization to generating instrument scripts and coordinating robotic lab work.Some approaches use mutual information to improve data efficiency; multimodal agents also support real-time interaction and error correction.Source: §4.1 · p. 6
4.2 Theory-Driven Scientific Paradigm
- In theory-driven science, FMs generate constrained hypotheses and support validation through symbolic reasoning, counterexample generation, and proof assistance.Examples include knowledge-graph-guided hypothesis formulation and systems coupling language models with symbolic solvers.Source: §4.2 · p. 6
4.3 Computation-Driven Scientific Paradigm
- In computation-driven science, FMs help formulate executable models and solve or invert equations using learned representations and function-space methods.Reported applications include discovering equations from diverse inputs, learning latent operators, accelerating weather prediction, and refining coarse solver outputs.Source: §4.3 · pp. 6–7
4.4 Data-Driven Scientific Paradigm
- In data-driven science, FMs extract knowledge from heterogeneous signals and use generative modeling for prediction and design.Examples range from DNA and chemistry representations to weather forecasts, protein structures, novel protein folds, and stable materials meeting specified constraints.Source: §4.4 · pp. 7–8
4 Foundation Model Integration Across Scientific Paradigms
- Across these paradigms, FMs are presented as mediators of hybrid workflows that combine methods traditionally used separately.The review characterizes this integration as enabled by general-purpose reasoning abilities, multimodal interfaces, and growing autonomy.Source: §4 · p. 5, §4.5 · p. 8
4.5 Cross-Paradigm Foundation Model Integration
- FMs connect experimental, theoretical, computational, and data-driven approaches in workflows spanning traditionally siloed methods.Examples include PROSE-FD and Latent Neural Operators, alongside Coscientist’s translation of research goals into robotic synthesis protocols that adapt to results.Source: §4.5 · p. 8
5 Risks of Emerging FM-Centered Scientific Paradigms and Future Direction
Risks intensify as FMs move from tools toward autonomous research agents. The section identifies key risk dimensions and outlines directions toward autonomous scientific discovery.
5.1 Risks of Emerging FM-driven Scientific Paradigms
- Training-data bias can steer scientific agendas toward dominant Western topics and away from underrepresented needs; proposed remedies include inclusive data, targeted fine-tuning, and fairness-aware evaluation.The paper gives global health as an example, contrasting well-studied diseases with issues such as schistosomiasis and child stunting in sub-Saharan Africa.Source: §5.1 · pp. 8–9
- Plausible but unverified FM-generated claims may misguide research or propagate when trusted prematurely; the paper calls for verification, human review, and provenance tracking.Examples include biomedical mechanisms without experimental grounding and physically invalid formulations.Source: §5.1 · p. 9
- Opaque reasoning, assumptions, or model versions can make FM-driven workflows difficult to replicate or validate; proposed safeguards include transparent logs and version-controlled checkpoints.The paper notes that a model-generated chemical synthesis pathway may lack interpretable derivations.Source: §5.1 · p. 9
- As FMs become collaborators or autonomous agents, authorship, accountability, and ethical conduct become more urgent, motivating governance that tracks contributions and downstream impacts.The paper raises questions about credit and responsibility when an FM generates a core hypothesis or experimental design.Source: §5.1 · p. 9
5.2 Future Directions: Toward Autonomous Scientific Discovery
- Embodied agents could connect language-based reasoning with physical perception and control to plan experiments and iteratively refine procedures, while facing robustness, safety, and interpretability challenges.The paper identifies laboratory robotics, automated instruments, and digital twins as deployment settings.Source: §5.2 · p. 9
- Closed-loop systems would let FMs formulate hypotheses, run experiments, analyze results, and update models from feedback; robustness to noise and alignment with scientific validity remain challenges.The paper cites reinforcement learning-based planning, planning-as-inference, and neuro-symbolic agents as current progress.Source: §5.2 · p. 9
- Continual learners could accumulate and refine cross-domain knowledge, but catastrophic forgetting, domain drift, and weak transfer across heterogeneous tasks and modalities remain obstacles.Suggested approaches include online adaptation, memory-augmented architectures, and modular lifelong learning frameworks.Source: §5.2 · p. 10
6 Conclusions
The conclusion presents a three-stage framework for the evolving role of FMs in science, from workflow support to autonomous discovery. It treats this transformation as emerging and calls for further study and responsible governance.
6 Conclusions
- The paper proposes meta-scientific integration, hybrid human-AI co-creation, and autonomous scientific discovery as stages in FM-enabled scientific paradigm evolution.Across classical paradigms, FMs increasingly act as epistemic agents rather than only tools.Source: §6 · p. 10
- The transformation remains emerging and raises questions about agency, authorship, and the nature of knowledge.The conclusion frames the possible shift as affecting who or what can do science.Source: §6 · p. 10
- The paper calls for rigorous exploration of FM capabilities, responsible governance mechanisms, and deeper theoretical understanding to guide FMs' role in science.It characterizes embracing the shift as potentially redefining how science is done.Source: §6 · p. 10