The formulation of research questions and hypotheses constitutes a fundamental aspect of scientific inquiry, providing a structured pathway for investigating phenomena and advancing knowledge. Research questions articulate specific gaps or uncertainties within a domain, whereas hypotheses propose tentative explanations or predictions amenable to empirical scrutiny. These elements ensure methodological rigour and facilitate the progression from observation to generalisation across disciplines.
Philosophical perspectives on hypothesis formulation emphasise the necessity of testability and refutation. Karl Popper advanced the criterion of falsifiability as a demarcation principle distinguishing scientific theories from non-scientific ones, positing that genuine scientific hypotheses must be incompatible with certain possible observations, thereby rendering them potentially (Thornton 2023). This approach, encapsulated in the method of conjecture and refutation, advocates the generation of bold conjectures followed by rigorous attempts at disconfirmation, rather than inductive verification, to mitigate biases and promote critical rationalism. Hypotheses, in this framework, derive from creative imagination and must exhibit high informative content, enabling risky predictions that invite empirical challenge. Such principles underscore the provisional nature of scientific knowledge, where corroborated theories approximate truthlikeness until supplanted by superior alternatives.
Traditional methods for developing research questions and hypotheses involve systematic steps, beginning with literature reviews to uncover unresolved issues. Researchers refine broad topics into focused questions that are specific, measurable, and feasible, often formulating hypotheses as null or alternative statements to support quantitative analysis. Nonetheless, these methods encounter limitations in handling voluminous data or interdisciplinary complexities, where human cognitive constraints may impede comprehensive exploration.
Artificial intelligence (AI) has profoundly transformed the formulation of research questions and hypotheses, leveraging computational capabilities to generate novel propositions, accelerate validation, and explore vast combinatorial spaces. Large language models (LLMs) drive this shift by synthesising extensive literature and datasets, identifying hidden patterns, and proposing conjectures with minimal human input (Huang et al. 2025). Comprehensive surveys outline LLM-based methods, including basic prompting for initial ideas, adversarial approaches to spur innovation, and multi-agent systems that mimic academic debate for refinement (Alkan et al. 2025). For example, frameworks like MOOSE-Chem adapt LLMs to specific domains, generating hypotheses from molecular data, while retrieval-augmented generation (RAG) reduces hallucinations by grounding outputs in reliable sources (Yang et al. 2024). Generative AI further automates extraction from big data using unsupervised learning, natural language processing, and knowledge graphs, uncovering associations in genomics and neuroscience, such as genetic-protein links or novel drug targets in conditions like pulmonary fibrosis, thereby shortening discovery timelines from years to months.
The POPPER framework deploys LLM agents to conduct sequential falsification experiments, testing hypotheses' implications with strict error controls across fields like biology and sociology, matching human expert performance while drastically cutting validation time. Autonomous verification tools evaluate LLM outputs in controlled environments, highlighting strengths in independent ideation despite logical gaps (Huang et al. 2025). Benchmarks such as HypoBench provide systematic assessments of LLM hypothesis generation, balancing novelty against coherence to inform improvements (Liu et al. 2025).
Innovative architectures support interactive and autonomous workflows. The IRIS system blends human oversight with AI for iterative hypothesis refinement via hybrid intelligence (Garikaparthi et al. 2025). In biomedicine, tools like GPT-4o craft hypotheses for challenges such as cardiotoxicity, while hybrid causal graphs and LLMs enable testable ideas in behavioural studies (Li et al. 2025). Autonomous agents like AI-Researcher manage end-to-end pipelines, from question formulation to evaluation, using Monte Carlo tree search for optimised prioritisation (Tang et al. 2025). Generative setups in self-driving laboratories refine hypotheses through feedback loops, simulating interventions and causal inference for counterfactual testing (Canty et al. 2025). AI also aids interdisciplinary tasks, such as drafting proposals with creativity evaluations. Such modalities offer key benefits, including enhanced efficiency in hypothesis exploration and broader access to scientific innovation, by boosting the odds of viable conjectures in limited settings. However, challenges include interpretability issues, bias inheritance from training data, and risks of eroding human critical thinking through over-reliance (Alkan et al. 2025). Ethical guidelines stress transparency in AI application to uphold reproducibility and accountability (Resnik and Hosseini 2024).
In conclusion, the integration of Karl Popper's falsifiability principle with AI-driven methodologies represents a significant advancement in formulating research questions and hypotheses, combining philosophical rigour with computational efficiency to generate and validate propositions more rapidly. AI systems, such as large language models and agentic frameworks like POPPER, enable the synthesis of vast datasets, uncovering novel insights in fields like biomedicine while automating validation through sequential testing. This synergy accelerates scientific discovery, and bridges gaps in traditional methods, particularly in complex, data-intensive domains. However, to maximise benefits, ethical considerations must address biases, interpretability, and reproducibility, ensuring human oversight complements AI capabilities.
References:
1. Alkan, Atilla Kaan, Shashwat Sourav, Maja Jablonska, Simone Astarita, Rishabh Chakrabarty, Nikhil Garuda, Pranav Khetarpal, et al. 2025. “A Survey on Hypothesis Generation for Scientific Discovery in the Era of Large Language Models.” arXiv preprint arXiv:2504.05496. ^ Back
2. Canty, Richard B., Jeffrey A. Bennett, Keith A. Brown, Tonio Buonassisi, Sergei V. Kalinin, John R. Kitchin, Benji Maruyama, et al. 2025. “Science Acceleration and Accessibility with Self-Driving Labs.” Nature Communications 16, no. 1: 3856. ^ Back
3. Garikaparthi, Aniketh, Manasi Patwardhan, Lovekesh Vig, and Arman Cohan. 2025. “Iris: Interactive Research Ideation System for Accelerating Scientific Discovery.” arXiv preprint arXiv:2504.16728. ^ Back
4. Huang, Kexin, Ying Jin, Ryan Li, Michael Y. Li, Emmanuel Candès, and Jure Leskovec. 2025. “Automated Hypothesis Validation with Agentic Sequential Falsifications.” arXiv preprint arXiv:2502.09858. ^ Back
5. Li, Yilan, Tianshu Gu, Chengyuan Yang, Minghui Li, Congyi Wang, Lan Yao, Weikuan Gu, and DianJun Sun. 2025. “AI-Assisted Hypothesis Generation to Address Challenges in Cardiotoxicity Research: Simulation Study Using ChatGPT With GPT-4o.” Journal of Medical Internet Research 27: e66161. ^ Back
6. Liu, Haokun, Sicong Huang, Jingyu Hu, Yangqiaoyu Zhou, and Chenhao Tan. 2025. “HypoBench: Towards Systematic and Principled Benchmarking for Hypothesis Generation.” arXiv preprint arXiv:2504.11524. ^ Back
7. Resnik, David B., and Mohammad Hosseini. 2024. “The Ethics of Using Artificial Intelligence in Scientific Research: New Guidance Needed for a New Tool.” AI and Ethics: 1–23. ^ Back
8. Tang, Jiabin, Lianghao Xia, Zhonghang Li, and Chao Huang. 2025. “AI-Researcher: Autonomous Scientific Innovation.” arXiv preprint arXiv:2505.18705. ^ Back
9. Thornton, Stephen. 2023. “Karl Popper.” In The Stanford Encyclopedia of Philosophy (Winter 2023 Edition), edited by Edward N. Zalta & Uri Nodelman. Available at: https://plato.stanford.edu/archives/win2023/entries/popper/ ^ Back
10. Yang, Zonglin, Wanhao Liu, Ben Gao, Tong Xie, Yuqiang Li, Wanli Ouyang, Soujanya Poria, Erik Cambria, and Dongzhan Zhou. 2024. “Moose-Chem: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses.” arXiv preprint arXiv:2410.07076. ^ Back
