Over the past decades, alongside advancements in natural language processing, significant attention has been paid to training models to automatically extract, understand, test, and generate hypotheses in open and scientific domains. However, interpretations of the theyrpmothesis for various natural language understanding (NLU) tasks have migrated from traditional definitions in the natural, social, and formal sciences. Even within NLU, we observe diferences defining hypotheses across literature. In this paper, we overview and delineate various definitions of hypothesis. Especially, we discern the nuances of definitions across recently published NLU tasks. We highlight the importance of well-structured and well-defined hypotheses, particularly as we move toward a machine-interpretable scholarly record.
The word “hypothesis” has been used variably with diferent meanings, over decades and centuries,
across the social, natural, and formal sciences—from its conceptual roots in ancient Greek philosophy
[
We argue that ambiguity and variability around the definition of hypotheses, which was once acceptable—even perhaps productive—is now a critical concern in light of natural language processing (NLP) and natural language understanding (NLU) tasks requiring quantitative operationalization of hypotheses, in particular, hypothesis extraction (detection/identification), verification, and generation. Emerging technical work in these fields often do not include explicit definitions of hypotheses, claims, or evidence, instead relyindegfacto on benchmark datasets to provide implicit definitions.
Following, we survey the definitions and operationalizations of hypotheses, focusing on research hypotheses engaged in the hypothesis mining literature. Research hypotheses are hypotheses designed for systematic investigation within a research framework. In this paper, we do not distinguish between research hypothesis and scientific hypothesis. In principle, these two terms have diferent scopes, but in practice, they are often used interchangeably in modern hypothesis mining papers. The related tasks are particularly in the areas of natural language inference (NLI), hypothesis extraction, scientific hypothesis evidencing and scientific claim verification, and scientific hypothesis generation. We highlight important Japan ∗The two authors made equal contributions to this paper. https://www.cs.odu.edu/~jwu(/J. Wu); https://www.rajtmajerlab.n(eSt./Rajtmajer)
CEUR
ceur-ws.org diferences and discuss the challenges these diferences impose on knowledge assembly and aggregation. Our work is motivated by the vision of a computable scholarly record—a verifiable and extensible knowledge base synthesizing computationally and data-enabled disco1v0e]r. ieTsh[is vision, we suggest, will be enabled by machine-readable hypotheses well-structured in predictable formats.
The wordhypothesis derives from Greek and means, literaallpyu,tting under or supposition. Ancient Greek philosophers used the term to describe a foundational assumption upon which to build out further reasoning. Plato uses the term in several of his dialogues with this intention–namely, as a claim accepted temporarily in order to explore its implications. Of particular interest to Plato was whether a hypothesis could support consistent and coherent conclu1s1io,1n2s][. Aristotle also engaged with the term. Aristotle viewed hypotheses more cautiously, being skeptical of relying on hypotheses without empirical verification. He delineated tentative assumptions from axioms or first principles, insisting that scientific knowledge must be based on demonstrated causes, not just assumed pre13m]i.ses [
During the scientific revolution of the 16th and 17th centuries, the concept of hypothesis evolved significantly. Galileo and Newton began using hypotheses as formal components of the scientific method, emphasizing the importance of testing through observation and exper1i4m, e1n5]t. [René Descartes also contributed to this shift, promoting skepticism and the formulation of testable prop1o6,si1t7i]o. ns [
By the 19th and 20th centuries, the hypothesis had become a central pillar in science. Prominent
philosophers of science Karl Popper and Thomas Kuhn centered hypotheses within the scientific process,
both agreeing that hypotheses must engage with empirical data in some way, i.e., should be testable,
observable1[
In the last decades, there has been a growing interest in hypothesis mining in scientific literature, mostly in the fields of NLP and NLU, but also in interdisciplinary fields between social science and AI. The exact definitions of hypotheses involved are not always provided in the context of research problems, and the specific forms and expressions vary across papers. Here, we categorize modern hypotheses into several types, which may deviate from traditional definitions2,3e]..g., [
Ideas as hypotheses. As defined in Kuhn & Hawkins [
a human is a kind of animal Question: Why do astronauts need oxygen in the backpacks of their spacesuits? an animal requires oxygen to breathe an astronaut is a kind of animal a vacuum does not contain oxygen space is a vacuum an astronaut requires oxygen
to breath spacesuit backpacks contain
oxygen there is no oxygen in space Answer: to help astronauts breathe in outer space
H: an astronaut requires the oxygen in a spacesuit backpack to breath generated content contains preliminary and broad notations intended to inspire further investigation, which is aligned with the concept of ideas. An example is shown in T1.able
Claims as hypotheses. The classical definition ocflaim is the conclusion or assertion that you want
your audience to accep2t6[]. Adopting this definition, in scientific literature, a claim can be defined as a
specific assertion reported as a finding of the paper. A paper can make more than one claim, and a claim
may contain one or multiple sentences. One definition of hypothesis is a claim that has not been tested
[27]. In Alipourfard et al2.8[], authors labelcalaim trace for each paper in their corpora, and each
trace contains four claims. Hypotheses and evidence are treated as two types of claims. In the recent
SciHyp dataset2[
Hypothesis-proposals. In recent works about hypothesis generation, models are built to generate not only a hypothesis but also a series of related sections such as its background, justification, and test procedures, resulting ipnroaposal-style document32[]. The hypothesis-proposal increases the transparency of hypothesis generation and provides a guide for testing. However, the specific format/sections of the proposal difer by model. An example3i2n] i[s shown in Table1.
Formal expressions. In early work, a research hypothesis is broken down into three dimensions, namely contexts, variables, and relations1h6i]p.sE[ach hypothesis is associated with a target variable and a set of independent variables, and relationships refer to the interactions between a given set of variables under a given context that produces the hypothesis. A hypothesis is then naturally expressed with asemantic tree in which the nodes represent variables and the edges represent relationships.
In more recent works in NLU, papers have expressed research hypotheses in various ways, depending
on the focal tasks. For example, in hypothesis generation tasks, the generated hypothesis may be
composed of multiple declarative statements, in which one serves as the main hypothesis and the others
provide additional context or details3(3s]eaen[d [25] in Table1). In the hypothesis evidencing task,
hypotheses can be written as questio3n0s],[which can be converted into hypotheses in declarative
form. A research hypothesis can also be decomposed as a question and an answer, e.g., the SciTail
dataset3[
Natural language inference (NLI, i.e., recognizing textual entailment (RTE)) involves assessing whether
a given textual premise entails or implies a given hypoth36e,si3s7[]. Most NLI datasets, such as SNLI
[38] and RTE-6 [39], are in open domains3[
Another scientific NLI datasetEisntailmentBank, built for multi-step scientific inference (Fig1u)r.e The task is to generate an entailment tree given a hypothesis. The tree shows a hierarchical supportive Question: Is there an association between social media use and bad mental health outcomes?
studies indicating an association Study1 Study2 Study3 studies indicating little or no association Study4 Study5
studies showing mixed evidence Study6 Study7 Study8
H1: There is an association between social media use and bad mental health out- comes.
H2: There is little or no association between social media use and bad mental health out- comes.
H3: There is little or no association between social media use and bad mental health out- comes. structure of claims toward a hypothesis. Other scientific NLI datasets include M4e0d]iaNnLdI [BioNLI [41] in the medical and biomedical domains, and e-SNLI-3V8E] [for visual scientific inference.
Here, the goal is to automatically identify hypotheses from a scientific document. Because hypotheses
can be viewed as claims prior to testing, the input, output, and methods for extracting hypotheses and
claims are similar. The input document can be an abstract42,,e.4g3.,] [or a full paper2[
Claim extraction may benefit from structured abstracts which contain, e.g., a dFeidnidciantgsed section (see Lancet45[]). Yet still, the specific sections difer across journals. For example4,2i]n, [ ifndings orproposed items are labeled as claims and one abstract may contain multiple claim1)s. (Table
It is worth noting that papers in computer science and several other domains often claim findings or contributions without explicitly stating hypothese2s5,]e..Wg.,e[consider these findings or contributionpsseudo-hypotheses.
Scientific hypothesis evidencing (SHE) 3[0] is the task of automatically identifying evidence from scientific publications in support of or refutation of a given hypothesis. This task is similar to another task called Scientific claim verification (SCV), both reflecting a model’s reasoning capability. The main diference is that in SHE, the hypotheses are usually high-level research questions (2F)i.gIunrSeCV datasets, the hypotheses are usually lower-level claims in specific contexts. However, certain cases are in between. Both tasks can be divided into two subtasks: identifying evidence candidates from a corpus of documents; and discerning the relationship between the hypothesis (claim) and an evidence candidate. Most research focuses on the second subtask, in which the relationships are classified into exclusive categories, nameSlUyPPORT, REFUTE, andNEI (not enough information) or their variants.
In Koneru et al.30[] authors build a dataset for the task of SHE using community-driven annotations
of studies in social sciences. The input is a hypothesis and an abstract (i.e., candidate evidence), and the
output is a label indicating whether the abstract entails, contradicts, or is inconclusive to the hypothesis.
In Wadden et al.3[
DiscoveryBench 4[
Output hypothesis: We hypothesize that integrating biomimetic materials with microfluidic chips will significantly enhance their heat transfer performance and biocompatibility, making them ideal for advanced biomedical applications.
Specifically, we propose that the lamellar structure of biomaterials, inspired by keratin scales, can be engineered into microfluidic chips using soft lithography techniques to improve their mechanical behavior and heat transfer eficiency under cyclic loading conditions.
Output other sections: outcome, mechanisms, design principles, unexpected properties, comparison, and novelty, and their expanded versions.
Input: Tweet pairs in the Tweet Popularity dataset [51].
Output: Tweets with named entities like people, places, or organizations tend to get more retweets by being more specific.
Input seed term: diverse relational edge embedding Input background: the task of converting a natural lanuage question into an executable sql query , known as text - to -sql, is an important branch of semantic parsing . the state - of - the -art graph -based encoder has been successfully used in this task but does not model the question syntax well.
Output: We propose a novel graph-based encoder that uses a diverse relational edge embeddings to model the question syntax.
Gen
Keywords (SciAgents [33])
Proposal Gen compared against the gold standard hypothesis (see an example in1T).aSbclieClaimHunt is a dataset recently built for SCV, in which a small amount of claims are manually extracted from the discussion and conclusion sections of research papers in computer science. Most claims are generated by LLMs. Of note, a fraction of claims are not self-contained and require reference to the context of the source paper.
The goal of scientific hypothesis generation is to automatically create new, testable scientific hypotheses or research ideas that identify novel relationships, phenomena, or gaps in existing kno2w4]l.edge [ Recent advancements in LLMs, e.g., Llam5a2][and GPT [53], ofer promise. Here too, existing literature is inconsistent with respect to task formulatioinnp,ui.te.a, ndoutput (see examples in Table1).
We identify four types of input for the hypothesis generation tasks:
(1) keywords, concise descriptions of the topics or central concepts of the model, such as in 2S5c]iMon [
and SciAgents3[
Likewise, we identify three types of output of hypothesis generation tasks: (1) traditional hypothesis, usually expressed as a single or multiple declarative sentence5s0,,e5.g8.],. [ (2) ideas, enriched hypotheses as shown in Secti2o.n2. For example, inSciMon, generated ideas may contain claims, methods, and objectives extracted from abstracts. (3) hypothesis proposals, comprehensive structured hypothesis description documents. For example, the output of SciAgent3s3[] is a document containing hypotheses, outcomes, mechanisms, design principles, unexpected properties, comparison, and novelty, each having its expanded version. AI co-sc3i2e]ntist [ also outputs a structured document but with diferent sections: introduction, recent findings, related research, rationale, specificity, experimental design, and validation. Si5e4t]raelq. u[est LLM agents to generate an “idea”, containing several components (e.g., problem, existing methods, motivation, proposed method, experiment plan), similar to a research proposal. Whereas, the Piflow fram5e5w] ork [ requires LLM agents to generate a “hypothesis structure” consisting of rationale, hypothesis, reiterate, and an experimental candidate.
Our work here intends to be more descriptive than prescriptive. We outline the various definitions and instances of hypotheses in existing scientific literature (and beyond). In particular, we focus on definitions of hypotheses and related concepts in recent work in NLU.
We hope that this work may raise awareness within the hypothesis mining community about standardization of corpus-level tools, e.g., knowledge graphs representing connections amongst interdisciplinary hypotheses, or hypothesis generation models across multiple domains. In lieu of standardization, the inclusion of explicit, clear definitions of hypotheses (formal, where possible) could improve alignment and assembly.
For more than two decades, many in the research community have advocated for open data, open materials, preregistration, and other best practices as central to thesevairsciohanbolef aand interpretable scholarly record. With recent technological advances, this vision is on the horizon. The ultimate goals of an interpretable scholarly record are: robust and eficient scientific progress; thoughtful allocation of community resources toward important open problems; and honest dialogue with public and policymakers. How–precisely–a queryable scholarly corpus comes together is an open question. Here, we suggest that dissemination of clear, consistent, well-specified machine-readable hypotheses, claims, and evidence are critical to this mission.
This project is partially supported by Open Philanthropy.
In accordance with the principles of responsible AI use, we disclose that generative AI was used solely for language editing and did not contribute to the scientific content or analysis presented in this work. research ideas?, CoRR abs/2409.06185 (2024). URLh:ttps://doi.org/10.48550/arXiv.2409.061.85 doi:10.48550/ARXIV.2409.06185. arXiv:2409.06185. [25] Q. Wang, D. Downey, H. Ji, T. Hope, SciMON: Scientific Inspiration Machines Optimized for Novelty, in: L. Ku, A. Martins, V. Srikumar (Eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, Association for Computational Linguistics, 2024, pp. 279–299. URL: https://doi.org/10.18653/v1/2024.acl-long..1d8oi:10.18653/V1/2024.ACL-LONG.18. [26] S. E. Toulmin, The uses of argument, Cambridge university press, 2003. [27] T. Heger, A. Algergawy, M. Brinner, J. M. Jeschke, B. König-Ries, D. Mietchen, S. Zarrieß, Natural language hypotheses in scientific papers and how to tame them: Suggested steps for formalizing complex scientific claims, in: Robust Argumentation Machines: First International Conference, RATIO 2024, Bielefeld, Germany, June 5–7, 2024, Proceedings, SpringerVerlag, Berlin, Heidelberg, 2024, p. 3–19. URhLt:tps://doi.org/10.1007/978-3-031-63536-6_.1 doi:10.1007/978-3-031-63536-6_1. [28] N. Alipourfard, B. Arendt, D. M. Benjamin, N. Benkler, M. Bishop, M. Burstein, M. Bush, J. Caverlee, Y. Chen, C. Clark, et al., Systematizing confidence in open research and evidence (score), SocArXiv (2021). [29] R. Vasu, C. Sarasua, A. Bernstein, Scihyp: A fine-grained dataset describing hypotheses and their components from scientific articles, in: International Semantic Web Conference, Springer, 2024, pp. 134–152. [30] S. D. Koneru, J. Wu, S. Rajtmajer, Can large language models discern evidence for scientific hypotheses? case studies in the social sciences, in: N. Calzolari, M. Kan, V. Hoste, A. Lenci, S. Sakti, N. Xue (Eds.), Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, LREC/COLING 2024, 20-25 May, 2024, Torino, Italy, ELRA and ICCL, 2024, pp. 2787–2797. URLh:ttps://aclanthology.org/2024.lrec-main..248 [31] D. Wadden, S. Lin, K. Lo, L. L. Wang, M. van Zuylen, A. Cohan, H. Hajishirzi, Fact or fiction: Verifying scientific claims, in: B. Webber, T. Cohn, Y. He, Y. Liu (Eds.), Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, Association for Computational Linguistics, 2020, pp. 7534–7550. URL: https://doi.org/10.18653/v1/2020.emnlp-main.6.0d9oi:10.18653/V1/2020.EMNLP-MAIN.609. [32] J. Gottweis, W. Weng, A. N. Daryin, T. Tu, A. Palepu, P. Sirkovic, A. Myaskovsky, F. Weissenberger, K. Rong, R. Tanno, K. Saab, D. Popovici, J. Blum, F. Zhang, K. Chou, A. Hassidim, B. Gokturk, A. Vahdat, P. Kohli, Y. Matias, A. Carroll, K. Kulkarni, N. Tomasev, Y. Guan, V. Dhillon, E. D. Vaishnav, B. Lee, T. R. D. Costa, J. R. Penadés, G. Peltz, Y. Xu, A. Pawlosky, A. Karthikesalingam, V. Natarajan, Towards an AI co-scientist, CoRR abs/2502.18864 (2025). UhtRtLp:s://doi.org/10. 48550/arXiv.2502.18864. doi:10.48550/ARXIV.2502.18864. arXiv:2502.18864. [33] A. Ghafarollahi, M. J. Buehler, Sciagents: Automating scientific discovery through bioinspired multi-agent intelligent graph reasoning, Advanced Materials 37 (2025) 2413523. URL: https://advanced.onlinelibrary.wiley.com/doi/ abs/10.1002/adma.202413523. doi:https://doi.org/10.1002/adma.202413523. arXiv:https://advanced.onlinelibrary.wiley.com/doi/pdf/10.1002/adma.202413523. [34] T. Khot, A. Sabharwal, P. Clark, SciTaiL: A Textual Entailment Dataset from Science Question Answering, Proceedings of the AAAI Conference on Artificial Intelligence 32 (2018)h.UttRpLs:: //ojs.aaai.org/index.php/AAAI/article/view/12.0d2o2i:10.1609/aaai.v32i1.12022. [35] B. Dalvi, P. Jansen, O. Tafjord, Z. Xie, H. Smith, L. Pipatanangkura, P. Clark, Explaining answers with entailment trees, in: M. Moens, X. Huang, L. Specia, S. W. Yih (Eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, Association for Computational Linguistics, 2021, pp. 7358–7370. URL:https://doi.org/10.18653/v1/2021.emnlp-main.5.8d5oi:10. 18653/V1/2021.EMNLP-MAIN.585. [36] I. Dagan, D. Roth, F. Zanzotto, M. Sammons, Recognizing textual entailment: Models and applications, Springer Nature, 2022. [37] S. Storks, Q. Gao, J. Y. Chai, Recent advances in natural language inference: A survey of benchmarks, resources, and approaches, arXiv preprint arXiv:1904.01172 (2019). [38] V. Do, O.-M. Camburu, Z. Akata, T. Lukasiewicz, e-snli-ve: Corrected visual-textual entailment with natural language explanations, arXiv preprint arXiv:2004.03744 (2020). [39] L. Bentivogli, P. Clark, I. Dagan, D. Giampiccolo, The fith pascal recognizing textual entailment challenge., TAC 7 (2009) 1. [40] C. Shivade, Mednli—a natural language inference dataset for the clinical domain, (No Title) (2017). [41] M. Bastan, M. Surdeanu, N. Balasubramanian, BioNLI: Generating a biomedical NLI dataset using lexico-semantic constraints for adversarial examples, in: Y. Goldberg, Z. Kozareva, Y. Zhang (Eds.), Findings of the Association for Computational Linguistics: EMNLP 2022, Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022, pp. 5093–5104.hUttRpLs:: //aclanthology.org/2022.findings-emnlp.3.7d4o/i:10.18653/v1/2022.findings-emnlp.374. [42] T. Achakulvisut, C. Bhagavatula, D. Acuna, K. Kording, Claim extraction in biomedical publications using deep discourse model and transfer learning, arXiv preprint arXiv:1907.00962 (2019). [43] X. Wei, M. R. U. Hoque, J. Wu, J. Li, Claimdistiller: Scientific claim extraction with supervised contrastive learning, in: Proceedings of Joint Workshop of the 4th Extraction and Evaluation of Knowledge Entities from Scientific Documents (EEKE2023) and the 3rd AI + Informetrics (All2023) co-located with the JCDL 2023, Santa Fe, New Mexico, United States, June 26 – June 30, 2023, 2023, pp. 3487–3496. URL: https://ceur-ws.org/Vol-3451/paper11.p.df [44] E. White, K. B. Cohen, L. Hunter, The CISP annotation schema uncovers hypotheses and explanations in full-text scientific journal articles, in: Proceedings of BioNLP 2011 Workshop, BioNLP ’11, Association for Computational Linguistics, USA, 2011, p. 134–135. [45] Lancet, Information for Authorhst,tps://www.thelancet.com/pb-assets/Lancet/authors/ tl-info-for-authors-1740074875577.p,d2f025. [46] A. Saakyan, T. Chakrabarty, S. Muresan, COVID-fact: Fact extraction and verification of real-world claims on COVID-19 pandemic, in: C. Zong, F. Xia, W. Li, R. Navigli (Eds.), Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Association for Computational Linguistics, Online, 2021, pp. 2116–2129. URhLt:tps://aclanthology.org/2021.acl-long..165/ doi:10.18653/v1/2021.acl-long.165. [47] M. Sarrouti, A. Ben Abacha, Y. Mrabet, D. Demner-Fushman, Evidence-based fact-checking of health-related claims, in: M.-F. Moens, X. Huang, L. Specia, S. W.-t. Yih (Eds.), Findings of the Association for Computational Linguistics: EMNLP 2021, Association for Computational Linguistics, Punta Cana, Dominican Republic, 2021, pp. 3499–3512. UhRtLt:ps://aclanthology.org/ 2021.findings-emnlp.297./doi:10.18653/v1/2021.findings-emnlp.297. [48] B. P. Majumder, H. Surana, D. Agarwal, B. D. Mishra, A. Meena, A. Prakhar, T. Vora, T. Khot, A. Sabharwal, P. Clark, Discoverybench: Towards data-driven discovery with large language models, in: The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025, OpenReview.net, 2025. URhL:ttps://openreview.net/forum?id= vyflgpwfJW. [49] K. BahadarKhan, A. A Khaliq, M. Shahid, A morphological hessian based approach for retinal blood vessels segmentation and denoising using region based otsu thresholding, PLOS ONE 11 (2016) 1–19. URL: https://doi.org/10.1371/journal.pone.01589.9d6oi:10.1371/journal.pone.0158996. [50] Y. Zhou, H. Liu, T. Srivastava, H. Mei, C. Tan, Hypothesis generation with large language models, in: L. Peled-Cohen, N. Calderon, S. Lissak, R. Reichart (Eds.), Proceedings of the 1st Workshop on NLP for Science (NLP4Science), Association for Computational Linguistics, Miami, FL, USA, 2024, pp. 117–139. URL: https://aclanthology.org/2024.nlp4science-.1d.1o0i/:10.18653/v1/2024. nlp4science-1.10. [51] C. Tan, L. Lee, B. Pang, The efect of wording on message propagation: Topic- and authorcontrolled natural experiments on Twitter, in: K. Toutanova, H. Wu (Eds.), Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, Baltimore, Maryland, 2014, pp. 175–185. URL: https://aclanthology.org/P14-10.1d7o/i:10.3115/v1/P14-1017. [52] H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, D. Bikel, L. Blecher, C. Canton-Ferrer, M. Chen, G. Cucurull, D. Esiobu, J. Fernandes, J. Fu, W. Fu, B. Fuller, C. Gao, V. Goswami, N. Goyal, A. Hartshorn, S. Hosseini, R. Hou, H. Inan, M. Kardas, V. Kerkez, M. Khabsa, I. Kloumann, A. Korenev, P. S. Koura, M. Lachaux, T. Lavril, J. Lee, D. Liskovich, Y. Lu, Y. Mao, X. Martinet, T. Mihaylov, P. Mishra, I. Molybog, Y. Nie, A. Poulton, J. Reizenstein, R. Rungta, K. Saladi, A. Schelten, R. Silva, E. M. Smith, R. Subramanian, X. E. Tan, B. Tang, R. Taylor, A. Williams, J. X. Kuan, P. Xu, Z. Yan, I. Zarov, Y. Zhang, A. Fan, M. Kambadur, S. Narang, A. Rodriguez, R. Stojnic, S. Edunov, T. Scialom, Llama 2: Open foundation and finetuned chat models, CoRR abs/2307.09288 (2023). URhLt:tps://doi.org/10.48550/arXiv.2307.092.88 doi:10.48550/ARXIV.2307.09288. arXiv:2307.09288. [53] T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, D. Amodei, Language models are few-shot learners, in: H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, H. Lin (Eds.), Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. URhLt:tps://proceedings.neurips.cc/paper/ 2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.h t.ml [54] C. Si, D. Yang, T. Hashimoto, Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers, in: The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025, OpenReview.net, 2025. UhRtLt:ps: //openreview.net/forum?id=M23dTGWC Z.y [55] Y. Pu, T. Lin, H. Chen, PiFlow: Principle-aware Scientific Discovery with Multi-Agent Collaboration, CoRR abs/2505.15047 (2025). URL:https://doi.org/10.48550/arXiv.2505.150.4d7oi:10.48550/ARXIV. 2505.15047. arXiv:2505.15047. [56] M. Chai, E. Herron, E. Cervantes, T. Ghosal, Exploring scientific hypothesis generation with mamba, in: L. Peled-Cohen, N. Calderon, S. Lissak, R. Reichart (Eds.), Proceedings of the 1st Workshop on NLP for Science (NLP4Science), Association for Computational Linguistics, Miami, FL, USA, 2024, pp. 197–207. URL: https://aclanthology.org/2024.nlp4science-.1d.1o7i/:10.18653/ v1/2024.nlp4science-1.17. [57] Z. Yang, X. Du, J. Li, J. Zheng, S. Poria, E. Cambria, Large language models for automated opendomain scientific hypotheses discovery, in: L.-W. Ku, A. Martins, V. Srikumar (Eds.), Findings of the Association for Computational Linguistics: ACL 2024, Association for Computational Linguistics, Bangkok, Thailand, 2024, pp. 13545–13565. URLh:ttps://aclanthology.org/2024.findings-acl..804/ doi:10.18653/v1/2024.findings-acl.804. [58] B. Qi, K. Zhang, H. Li, K. Tian, S. Zeng, Z.-R. Chen, B. Zhou, Large language models are zero shot hypothesis proposers, arXiv preprint arXiv:2311.05965 (2023).