Skip to main content
NCBI home page
Frontiers in Artificial Intelligence logoLink to Frontiers in Artificial Intelligence
. 2025 Oct 9;8:1664303. doi: 10.3389/frai.2025.1664303

There are significant differences among artificial intelligence large language models when answering scientific questions

Francisco Javier Álvarez-Martínez 1, Luis Esteban 2, Lucas Frungillo 3, Estefanía Butassi 4, Alessandro Zambon 5, María Herranz-López 1, Mario Aranda 6, Federica Pollastro 7, Anne Sylvie Tixier 8, Jose V Garcia-Perez 9, David Arráez-Román 10, Andrew Ross 11, Pedro Mena 12, Ru Angelie Edrada-Ebel 13, James Lyng 14, Vicente Micol 1, Fernando Borrás-Rocher 15, Enrique Barrajón-Catalán 1,*
PMCID: PMC12547693  PMID: 41141903

Abstract

Introduction

This study investigates the efficacy of large language models (LLMs) for generating accurate scientific responses through a comparative evaluation of five prominent free models: Claude 3.5 Sonnet, Gemini, ChatGPT 4o, Mistral Large 2, and Llama 3.1 70B.

Methods

Sixteen expert scientific reviewers assessed these models in terms of depth, accuracy, relevance, and clarity.

Results

Claude 3.5 Sonnet emerged as the highest scoring model, followed by Gemini, with notable variability among the other models. Additionally, retrieval-augmented generation (RAG) techniques were applied to improve LLM performance, and prompts were refined to improve answers. The results indicate that although LLMs such as Claude 3.5 Sonnet have potential for scientific tasks, other models may require more development or additional prompt engineering to reach comparable accuracy. Reviewers’ perceptions of artificial intelligence (AI) utility and trustworthiness showed a positive shift after evaluation. However, ethical concerns, particularly with respect to transparency and disclosure, remained consistent.

Discussion

The study highlights the need for structured frameworks for evaluating LLMs and ethical considerations essential for responsible AI integration in scientific research. These findings should be interpreted with caution, as the limited sample size and domain-specific focus of the exam questions restrict the generalizability of the results.

Keywords: large language models, artificial intelligence, scientific evaluation, prompt engineering, retrieval-augmented generation

Graphical abstract

A scatter plot categorizes partners based on effectiveness and relevance. Areas outlined in pink highlight groups: "Unforeseen/Eventual", "Ineffective and Less Strategic", "Occasional", "Strong and Effective", "Relevant", and "Preferred/Diamond" partners. Data points are colored and labeled from one to six.

1. Introduction

Rapid advancements in artificial intelligence (AI), particularly in large language models (LLMs), are transforming the landscape of scientific research by streamlining knowledge discovery, data analysis, and hypothesis generation (Borger et al., 2023). LLMs are highly sophisticated models trained on vast and diverse text corpora, with architectures that enable them to process and generate human-like language (Rudolph et al., 2023). Recent developments have produced state-of-the-art models capable of deriving insights from complex, domain-specific data with remarkable depth and accuracy (Liu et al., 2024). However, while their potential in scientific applications is evident, a critical question remains: how do these models compare in their ability to accurately retrieve, interpret, and synthesize specialized information across different fields? Addressing this question is essential for understanding their reliability as tools for scientific inquiries.

Several recent studies have attempted to benchmark LLMs in scientific or technical contexts, often focusing on single models or narrow applications such as medical question answering, patient communication, or domain-specific tasks (Shool et al., 2023; Jahan et al., 2024; Li et al., 2024; Wang et al., 2024). These contributions provide valuable insights but generally lack systematic cross-model comparison combined with domain-specific retrieval augmentation and expert-based evaluation. By integrating these components, our study addresses this gap and proposes a reproducible benchmarking framework that expands beyond existing work.

This study compares leading LLMs systematically to determine which LLMs are best suited to answer complex scientific questions. By assessing the depth, accuracy, relevance, and clarity of each model, we establish a benchmark for its effectiveness in scientific research, revealing both its current strengths and limitations. Given the diverse architectures, training data, and design approaches across LLMs, this comparative analysis offers valuable insights for researchers and practitioners in various fields. A secondary aim of this study is to develop an optimized workflow to improve LLM performance in scientific contexts via retrieval-augmented generation (RAG) (Li and Ouyang, 2024) techniques for improved prompt generation. By expanding initial queries with targeted domain-specific information, RAG enables models to retrieve more relevant data. Our methodology, implemented through Jupyter notebooks (fully and freely available), refines question prompts with specific scientific literature, increasing the precision of LLM outputs and offering a practical framework for researchers to maximize these tools in their work.

This research contributes to the growing integration of AI in scientific inquiry and focuses on identifying which LLMs are best equipped for the complex interdisciplinary synthesis needed to drive scientific discovery. Our work stands out by employing a novel, objective, and quantifiable methodology that enables reliable comparisons of LLMs. The insights gained in this study will guide researchers in effectively leveraging AI for their needs, identify the strengths and limitations of prominent LLMs, and help chart future improvements toward a more reliable and powerful role for AI in scientific research.

2. Materials and methods

2.1. Selection of LLMs

Five state-of-the-art LLMs were selected for this study on the basis of their prominence and broad availability. The most advanced open-access LLMs available were used during August 2024. These models included Claude 3.5 Sonnet (Anthropic), Gemini (Google), GPT-4o (Open AI), Large 2 (Mistral AI) and Llama 3.1 70B (Meta). Each LLM was tested for its ability to answer a set of complex scientific questions pertaining to the extraction of bioactive compounds from agricultural byproducts.

2.2. Question set design

An exam composed of 10 scientifically rigorous questions was designed by the authors to assess the ability of each model to understand and synthesize knowledge. The questions listed below cover a wide range of key topics in the fields of agricultural byproducts, green extraction and sustainability:

  1. What are agricultural byproducts, and why are they significant in the context of compound extraction?

  2. What are the main advantages of using agricultural byproducts for compound extraction compared with traditional raw materials?

  3. Describe the general process of extracting and/or purifying bioactive compounds from agricultural byproducts.

  4. Explain the role of solvents in the extraction process and name at least three commonly used. Include at least one solvent recognized as novel or green.

  5. Identify and describe three common methods used for extracting compounds from agricultural byproducts.

  6. Discuss the environmental and economic impact of extracting compounds from agricultural byproducts.

  7. What are some of the challenges associated with the extraction of compounds from agricultural byproducts?

  8. Explain how the quality and composition of agricultural byproducts can affect the extraction process and yield of compounds.

  9. Describe a case study in which a specific compound was successfully extracted from an agricultural byproduct.

  10. What future trends or advancements do you foresee in the field of compound extraction from agricultural byproducts?

2.3. Retrieval-augmented generation (RAG) strategy

To obtain reliable contextual scientific information, a search using the terms “Extraction AND Agricultural AND Byproduct” was performed in the Scopus database on July 12, 2024, and 306 articles were obtained. The information from these articles, including their abstracts, was downloaded in a CSV file. The information from this file was used by a Jupyter notebook in Google Colab to serve as context for the scientific questions in the exams. Before the questions were answered, each LLM was tasked with performing a query expansion to refine the search and retrieval of scientific abstracts used as context before answer generation. The expanded queries allowed the models to select a more targeted set of relevant documents from scientific databases, thereby increasing the relevance and accuracy of their answers. By using embeddings after the query expansion was performed using each of the different LLMs, the article abstracts were selected to answer each of the questions. Using this information, “superprompts” were generated, including the specific context, the scientific question and clear instructions to answer each of the questions on each exam with each LLM. A complete set of prompts, questions, answers, databases, notebooks, and code is openly available in the OSF repository (see Data Availability Statement).

2.4. Answer generation

Once the query expansion and retrieval were complete, each LLM was instructed to answer the 10 exam questions one by one on the basis of the generated “superprompts.” Each question was asked with its corresponding “superprompt” in a new chat to minimize possible interference. The answers were anonymized and formatted uniformly for subsequent evaluation by human experts. Exams were coded as Exam 1, Exam 2, Exam 3, Exam 4 and Exam 5. Only the first author of this work, who did not evaluate any exams, knew which LLM had solved each exam. The order of the exams (1 to 5) was chosen by random number generation. A diagram of the process from obtaining the articles for the scientific context to obtaining the answers for each LLM is shown in Figure 1. All LLMs were accessed through their free public interfaces, which did not allow modification of hyper-parameters such as temperature or maximum tokens. As a result, all experiments were performed using the default configurations of each platform, which we recognize as a methodological limitation.

Figure 1.

Flowchart illustrating a process with eight labeled boxes: "Topic Selection" leads to "Exam & Rubric Creation," both of which direct to "Review and Score." "Selection of Reviewers" goes to "Pre-test," and both connect to "Review and Score." This then leads to "Post-test" and "Comparison." "Comparison Among LLMs" is linked from "Review and Score."

Open in a new tab

Study workflow. Gray boxes represent the stages of exam creation and evaluation, while blue boxes depict the steps undertaken by the co-authors serving as reviewers. The “Review and Score” box, colored both gray and blue, signifies its inclusion in both pipelines.

2.5. Expert evaluation panel

A panel of 16 human experts from 7 different countries and 14 different institutions (Supplementary Table 1) was assembled to evaluate the answers generated by the LLMs. These experts were selected on the basis of their research expertise in fields related to agricultural byproduct valorization, bioactive compound extraction, and green chemistry. The panel members have h-indices between 10 and 58, with a mean value of 35.06 and a median of 38.

All responses were anonymized and uniformly formatted before scoring, and evaluators were blinded to which LLM had generated each set of answers. Only the non-evaluating authors, who designed the exam and prompts but did not participate in the evaluation, knew the assignment of models to exams. A standardized rubric defining the criteria of accuracy, depth, relevance, and clarity was provided to all evaluators in advance. No formal calibration sessions were conducted, as the aim was to preserve the evaluators’ natural heterogeneity and thereby reflect the diversity of perspectives present in the broader scientific community. The panel itself was diverse in nationality, institutional background, age, gender and experience, which further strengthens the representativeness of the evaluation.

2.6. Evaluation criteria

The anonymized answers from each LLM were scored by the panel using a scale of 0 to 10, with a single standardized rubric containing the following evaluation criteria: accuracy, i.e., how correct and factually precise the answer is, consistent with the current state of scientific knowledge; depth, i.e., the level of detail and comprehensiveness provided in the explanation; relevance, i.e., how well the answer addressed the specific question; and clarity, i.e., the ease of understanding and logical structure of the response. The rubric was co-designed by the authors to ensure consistency across evaluations, with clear definitions for accuracy, depth, relevance, and clarity provided to all reviewers in advance. Each reviewer applied the same standardized criteria independently. Statistical analysis was performed using the Friedman test, which is appropriate for non-parametric repeated measures (Shi et al., 2024), to account for the fact that all LLMs were evaluated by the same panel of experts. This approach ensured methodological rigor despite the limited sample size.

2.7. Pre- and postevaluation questionnaires

To gather additional insights into expert perceptions, two sets of questionnaires were administered. A preevaluation questionnaire in which experts were asked about their familiarity with AI and LLMs, their expectations for AI performance in scientific applications, and any preconceptions they might have regarding the utility of AI in research. A postevaluation questionnaire in which, after reviewing the AI-generated answers, experts were surveyed on their perceptions of the quality of the responses and their views on the potential for LLMs to assist in scientific research.

2.8. Statistical analysis

The scores from the expert evaluations were compiled and averaged for each LLM. The results were then analyzed using the Friedman test to identify any significant differences in performance across the LLMs.

3. Results

To ensure a rigorous and precise evaluation of the various LLMs, a topic within the authors’ expertise was selected. An exam comprising 10 questions (see the Supplementary material) and a rubric for assessment were subsequently developed (gray boxes in Figure 1). Concurrently (blue boxes in Figure 1), a panel of 16 coauthors was assembled, and an initial survey (pretest) was administered to gage their prior experience with AI and LLM models. These panelists then proceeded to evaluate and score the responses generated by the different LLMs. After the evaluation, the coauthors who served as reviewers completed a second survey (posttest) that reflects their experiences. A comparative analysis was conducted between the scores and responses of the LLMs and the reviewers’ answers before and after the evaluation.

3.1. LLMs expert panel comparison

After the review and scoring process conducted by the panel of reviewers, the average scores and standard deviations achieved by each LLM were analyzed to assess their efficacy in generating scientifically accurate and relevant responses. All coauthors who served as reviewers evaluated the responses using a blind approach to mitigate any potential bias stemming from their prior experience or knowledge of LLMs.

To promote open access science, the five most prominent and free-access models were tested: Llama 3.1 70B, Gemini, Claude 3.5 Sonnet, ChatGPT 4o, and Mistral Large 2. Each model was blindly evaluated by a panel of 16 experts on the basis of its ability to provide accurate, comprehensive answers to an exam composed of a set of complex scientific questions regarding green chemistry and sustainability.

Claude 3.5 Sonnet achieved the highest average score, with a mean rating of 8.41 ± 1.53 (out of 10). Gemini followed with a mean score of 7.73 ± 1.50, reflecting a solid performance with relatively low variability. ChatGPT 4o also demonstrated competitive results, with an average score of 7.60 ± 1.55, showing its ability to produce reliable responses across various queries. Llama 3.1 70B and Mistral Large 2 received slightly lower average scores of 7.36 ± 1.51 and 7.15 ± 1.66, respectively. Mistral Large 2 had the highest standard deviation, which indicates slightly greater fluctuation in performance. There were significant differences among the grades obtained by the different LLMs (Table 1). Claude 3.5 Sonnet had a significantly better average grade than Llama 3.1 70 B (p < 0.01), Mistral Large 2 (p < 0.01) and ChatGPT 4o (p < 0.05); however, the average grade was not significantly different from that of Gemini (p = 0.38). A heatmap with comparisons of the statistical significance among the grades obtained by each LLM is shown in Supplementary Figure 1.

Table 1.

Average grades and standard deviations obtained by each LLM after evaluation by expert scientific reviewers.

LLM Mean Standard deviation
Llama 3.1 70B 7.15b 1.51
Gemini 7.73a,b 1.50
Claude 3.5 Sonnet 8.41a 1.53
ChatGPT 4o 7.60b 1.55
Mistral Large 2 7.36b 1.66
Open in a new tab

Different letters (a, b) indicate significant differences among LLMs on the basis of the Friedman test (p < 0.05).

These findings suggest that, while all the tested LLMs demonstrated competence in addressing scientific queries, Claude 3.5 Sonnet and Gemini outperformed the other models in terms of overall accuracy and reliability, with Claude 3.5 Sonnet emerging as the top performer. This comparative analysis provides valuable insights into the strengths and limitations of each model and guides researchers in selecting the most suitable LLM for applications that demand high-quality scientific outputs.

To improve visualization of the differences among the LLMs, the results were represented as a histogram (Figure 2A) and as a kernel density estimate plot for continuous probability density (Figure 2B). Claude 3.5 Sonnet (in green) stands out from the other four LLMs.

Figure 2.

Panel A shows a stacked histogram with six AI models, each represented by different colors, plotting grades on the x-axis and count on the y-axis. Panel B presents a line graph with the same models, showing grade density on the y-axis with a range of curves peaking around grades six to ten.

Open in a new tab

(A) Histogram showing the reviewers’ grades for the exam conducted by each LLM. (B) Kernel density estimate plot for LLM grades probability density.

The average scores obtained for each specific exam question for each LLM were analyzed (Figure 3). All the models exhibited a circular pattern, which suggests consistent performance across all the questions, with minor variations affecting the final score. However, a drastic decrease in the ChatGPT 4o grade can be clearly observed in question 9: “Describe a case study where a specific compound was successfully extracted from an agricultural byproduct.” In this case, the LLM focused on summarizing key points from the related literature but failed to provide a concrete example as it was requested. On the other hand, the superior performance of Claude 3.5 Sonnet on question 8, “Explain how the quality and composition of agricultural byproducts can affect the extraction process and the yield of compounds,” likely results from its ability to analyze and contextualize scientific information in detail. This model has excelled in identifying and articulating the specific relationships between the quality and composition of agricultural byproducts and their influence on extraction efficiency and yield. Its approach might reflect better integration of key concepts, a clearer structure in presenting cause–effect relationships, and a focus on critical variables, resulting in a more insightful and precise answer.

Figure 3.

Radar chart comparing five models across ten questions. Models are Llama 3.1 70B (blue), Gemini (orange), Claude 3.5 Sonnet (green), ChatGPT 40 (red), and Mistral Large 2 (purple). Each line shows performance across questions one to ten, with varying values up to ten.

Open in a new tab

Radial chart with the average scores of each LLM for each specific question in the exams evaluated by the panel of reviewers.

3.2. Evaluation questionnaires

In order to complement the performance evaluation of the LLMs, we also analyzed how reviewers’ perceptions of AI evolved before and after the assessment. While methodologically distinct, these two components are presented together in order to capture not only how the models performed but also how their use was perceived by domain experts.

To assess the evaluators’ baseline perspectives on AI tools before the LLM-generated scientific responses were reviewed, a preevaluation questionnaire was administered to the 16 reviewers. The questionnaire included five questions concerning their current AI usage, future intentions to use AI, perceived usefulness, trustworthiness, and views on ethical considerations. The responses provided insight into the prevailing attitudes toward AI applications in scientific research. After the LLM-generated responses were evaluated, a postevaluation questionnaire was administered to the same 16 reviewers. These questions were aimed at capturing shifts in perspectives, which enabled a comparison with the preevaluation questionnaire results and offered insight into whether and how direct engagement with AI-involved tasks influenced their views.

In the preevaluation questionnaire, the reviewers expressed mixed intentions regarding short-term AI use: 37.50% were unsure, 25.00% intended to start using AI, and 37.50% were already using AI. Postevaluation, intentions leaned toward greater adoption. Only 18.75% were still uncertain, and the proportion of those already using AI or planning to increase their AI usage increased to 81.25%. Notably, 31.25% indicated that they would use more AI tools than they currently do, which reflected a significant postevaluation shift toward the integration of AI in daily scientific work (Figure 4). This finding suggests that exposure to AI’s potential through the evaluation task heightened awareness of its applicability, fostering a positive shift in adoption intentions.

Figure 4.

Bar graph compares trust in AI scientific tools before and after a review. Pre-review: 12.5% very trustworthy, 50% neutral, 18.75% trustworthy, 12.5% very untrustworthy. Post-review: 12.5% very trustworthy, 50% neutral, 18.75% slightly trustworthy, 6.25% slightly untrustworthy.

Open in a new tab

Shift in AI tools adoption intentions among reviewers before and after evaluation.

The preevaluation survey revealed that 31.25% of the respondents rated AI as being minimally useful (rating of 2 out of 5), and only 12.50% considered it extremely useful (rating of 5 out of 5). After the evaluation task, perceptions of the utility of AI improved: 43.75% rated it moderately useful (rating of 4 out of 5), and another 25.00% found it extremely useful (rating of 5 out of 5). This upward trend signifies that interaction with AI-based models highlighted tangible benefits for some reviewers, who became more convinced of its relevance to scientific tasks (Figure 5), despite the remaining skepticism among 25.00% who still rated the usefulness of AI at a low level (rating of 2 out of 5).

Figure 5.

Bar charts display perceptions of AI tools' usefulness in scientific work before and after a review. Pre-review: 31.25% not at all useful, 25% slightly useful, 31.25% moderately useful, 12.5% extremely useful. Post-review: 25% not useful, 6.25% slightly useful, 43.75% moderately useful, 25% extremely useful.

Open in a new tab

Changes in reviewers’ perceptions of AI utility before and after evaluation.

Half of the reviewers (50.00%) reported a moderate shift in their views (rating 3 out of 5) postevaluation, whereas only 18.75% claimed to have experienced no change in perspective (rating of 1 out of 5). Moreover, 6.25% noted a significant positive shift (rating of 5 out of 5). The majority noted some level of change, which indicates that hands-on experience with AI output influenced perceptions, even if most adjustments were modest. The results suggested that direct engagement with AI may gradually influence scientific opinions, which encourages openness to AI integration over time.

For those whose opinions shifted, the change tended to be positive or neutral. Specifically, 37.50% expressed a neutral stance (rating of 3 out of 5), 37.50% viewed AI more favorably postevaluation (rating of 4 out of 5), and 18.75% noted a distinct positive shift (rating of 5 out of 5). Only one reviewer (6.25%) reported a slightly negative change (Figure 6). This distribution indicates that interaction with AI tools in a controlled scientific context generally supports positive adjustments in perception, which potentially alleviates initial doubts or misconceptions.

Figure 6.

Survey results are shown in two horizontal bar charts about changes in opinion on AI tools after evaluation. The first chart indicates how likely people changed their views, with the largest group at fifty percent having a neutral stance. The second chart measures sentiment change: thirty-seven point five percent remained neutral, while another thirty-seven point five percent had a positive change. Percentage labels and color keys explain the scale of opinion from negative to positive.

Open in a new tab

Shifts in reviewers’ perspectives on AI following the evaluation.

In terms of trust, the preevaluation responses were mixed, with only 18.75% of the respondents rating AI tools as highly trustworthy (rating of 4 or 5 out of 5). After the evaluation exercise, however, the trust ratings improved (Figure 7): 50.00% of the participants rated AI as fairly trustworthy (rating of 4 out of 5), and 12.50% rated it as very trustworthy (rating of 5 out of 5). These shifts indicate that reviewers became more comfortable with the reliability of AI tools after they observed their capabilities firsthand, although 18.75% still expressed low levels of trust (rating of 2 or 3 out of 5). This result points to a gradual building of confidence that may increase with further familiarity and reliability demonstrations.

Figure 7.

Survey results comparing pre- and post-views on using AI tools for scientific or bibliographic purposes in the short term. Pre-survey: 37.5% not using, 25% unsure, 37.5% will use. Post-survey: 18.75% not using, 18.75% unsure, 31.25% will start, 31.25% will use more.

Open in a new tab

Changes in reviewers’ trust in AI tools before and after evaluation.

When revisiting ethical considerations, reviewers largely upheld the need for AI usage disclosure, with 50.00% agreeing that authors should specify any AI usage in scientific communications, a percentage that is very close to the preevaluation percentage (54.17%). Moreover, 33.33% supported regulations on AI use, a perspective that remained stable compared with preevaluation views. Notably, 12.50% endorsed unrestricted AI usage, identical to initial responses, and one reviewer (4.17%) expressed uncertainty. This finding suggests that while positive shifts occurred regarding AI utility and trustworthiness, ethical concerns remained steady, with most reviewers advocating for transparency and regulatory frameworks (Figure 8).

Figure 8.

Bar charts comparing opinions on the ethical use of AI tools for scientific/bibliographic purposes, before and after a review. Pre-review shows 12.50% for prohibition, 54.17% for stating AI use, and 33.33% for regulation. Post-review shows unchanged prohibition, decreased AI use statements to 50.00%, and unchanged regulation percentage with an additional 4.17% feeling unqualified to answer. Legend clarifies bar colors.

Open in a new tab

Reviewers’ perspectives on ethical considerations for AI usage before and after evaluating.

4. Discussion

This study presents a novel contribution by rigorously comparing the performance of the most popular free LLMs in responding to complex scientific questions, as assessed by a diverse panel of international experts. This approach has not been extensively applied, which positions this work as an important benchmark for assessing LLM effectiveness (Figure 9).

Figure 9.

Flowchart depicting a process with several stages: Research questions, Article retrieval, Retrieval-Augmented Generation, "Superprompt" generation, Embedding, Query expansion, LLM inquiry, LLM answer, and Expert evaluation. Icons illustrate each step, with arrows indicating sequence.

Open in a new tab

Diagram of the process of preparing information and answering each of the exam questions.

A key limitation of this study is the limited sample size of exam questions, which inevitably constrains the scope of the conclusions. Although Claude 3.5 Sonnet achieved the highest score in our evaluation, this result should not be assumed to hold across other domains or scientific contexts. Indeed, our own findings support the reviewer’s observation that LLM performance can vary substantially depending on the subject area. For this reason, the outcomes of the present work should be regarded as preliminary and domain-specific. Future studies should increase both the number of questions and the diversity of disciplines in order to provide more robust validation.

Previous research has shown that LLMs can produce informative and often high-quality content across medical and health-related topics; however, these studies have typically focused on single LLMs or specific applications, such as generating patient information leaflets or assisting with medical decision-making (Pompili et al., 2024; Saibene et al., 2024; Sharifi et al., 2024; Goel et al., 2023). In a similar vein, recent studies have highlighted the potential of LLMs for performing complex tasks in medicine, such as health coaching and monitoring of chemotherapy-induced toxicity (Saibene et al., 2024; Ruiz Sarrias et al., 2024). However, these studies often rely on specific question types or isolated case studies, with limited cross-comparisons among LLMs. In contrast, our study addresses a broader spectrum of scientific queries in the context of food technology, agricultural byproducts, green extraction and sustainability, spanning different levels of complexity and contextual depth, which are fundamental to evaluating robustness of each model in real-world scientific applications.

A key strength of this study is its methodological design, which uses a RAG technique to optimize query relevance, a method that has been shown to increase LLM outputs by incorporating domain-specific knowledge (Cheng et al., 2023; Deiner et al., 2024). This augmentation aligns with recent calls in the literature for integrating specialized data sources to improve LLM performance on expert-level tasks, particularly in areas such as oncology and mental health, where nuanced information is essential (Ruiz Sarrias et al., 2024; Elyoseph and Levkovich, 2024). Additionally, our inclusion of a large and diverse expert panel for response evaluation adds depth to our findings. Unlike studies that rely solely on quantitative metrics, our approach captures qualitative insights regarding response accuracy, readability, and helpfulness—dimensions that are critical for scientific applicability but are often underexplored (Pompili et al., 2024; Kim and Kim, 2024).

Previous works have highlighted certain limitations in LLM-generated responses, such as difficulty in conveying specific case examples or maintaining accuracy in highly specialized topics (Mu et al., 2024; Kostic et al., 2024; Chen, 2024). Our results corroborate these findings, especially for ChatGPT 4o, which underperformed on a question requiring detailed case-based reasoning. This outcome suggests that LLMs, while valuable, may still struggle with domain-specific precision unless complemented by further prompt optimization or RAG techniques. Nonetheless, the high performance of Claude 3.5 Sonnet in delivering accurate and contextually relevant responses points to the potential of next-generation LLMs to meet the demands of complex scientific discourse, potentially paving the way for AI to become a reliable collaborator in scientific research (Deiner et al., 2024).

In this study we chose to report and compare the average scores assigned to each LLM by the expert panel. This approach was selected as it provides an intuitive and interpretable measure of overall performance, while also allowing statistical comparison across models. We acknowledge that complementary descriptors, such as medians or distributions of scores, could provide additional insights and should be considered in future studies.

An important methodological consideration is that all evaluators were also co-authors. However, we do not regard this as a conflict of interest, since the evaluation process was fully blinded: none of the evaluators knew which LLM generated each answer, and only the first authors who did not serve as evaluators had access to this information. Several safeguards were in place, including anonymization of all responses and independent scoring without discussion among panel members. Furthermore, while all evaluators followed the same standardized rubric, no calibration beyond this rubric was imposed, in order to reflect the heterogeneity of scientific judgment. This approach, combined with the diversity of our panel (in terms of country, institution, age, and gender), ensures that the results reflect a realistic range of expert perspectives rather than artificially homogenized assessments.

By positioning this study within the broader landscape of LLM research, our findings provide a comprehensive view of current LLM capabilities and highlight the importance of continuous model refinement. Future research should build on this approach by exploring more interactive and adaptive models, particularly those capable of real-time adjustments based on human feedback, to better meet the evolving needs of scientific inquiry (Cheng et al., 2023; Kim and Kim, 2024). As AI continues to integrate into research workflows, establishing reliable benchmarks for LLM performance and ethical guidelines for their use will be essential in safeguarding the quality and integrity of scientific output (Elyoseph and Levkovich, 2024).

While the exploratory nature and domain specificity of this study may resemble certain aspects of a community case study, we believe the systematic benchmarking, statistical analysis, and integration of expert perspectives clearly situate this work within the scope of original research. By combining quantitative evaluation with qualitative insights, the study provides a rigorous and reproducible framework that contributes to the broader field of LLM assessment in scientific contexts.

The performance disparities among LLMs, as evidenced in our comparative analysis, highlight a critical issue: inconsistencies in AI-generated outputs can result in divergent interpretations when LLMs are relied upon for scientific purposes (Qu and Wang, 2024; Wu et al., 2024). This variability not only affects the perceived reliability of LLMs but also raises questions regarding their interpretability and transparency in research contexts. Our results stress the need for transparency and explainability in AI tools used in scientific research, because these factors are essential for building trust among researchers and maintaining rigorous scientific standards.

The ethical considerations expressed by our panel of reviewers both pre- and postevaluation emphasize the importance of disclosure and regulatory oversight when LLMs are integrated into research workflows. Transparency regarding AI involvement in scientific analysis, as supported by more than half of our reviewers, is crucial for preserving research integrity and mitigating potential biases introduced by AI systems. These findings resonate with the growing consensus in the AI ethics literature, which advocates for disclosure practices that inform readers of AI contributions (Perkins, 2023; Hosseini et al., 2023; Kiseleva et al., 2022; Chen et al., 2022), thereby fostering accountability and trust.

The shift in reviewer perspectives after direct interaction with LLM-generated outputs suggests that familiarity and hands-on experience with AI tools may alleviate initial skepticism. Nonetheless, the consistent emphasis on ethical disclosure implies a foundational need for clear regulatory guidelines to govern AI use in scientific research. Establishing these frameworks would provide researchers with a structured pathway for ethically incorporating AI, thereby aligning with the broader goals of transparency, reliability, and societal trust in AI-driven research.

Future research should expand this benchmarking approach by incorporating a broader array of LLMs and scientific disciplines, thereby providing a comprehensive understanding of LLM utility across fields. Additionally, advancements in LLM interpretability and transparency tools, such as model auditing and explainable AI (XAI) techniques (Ali et al., 2023), could address several ethical concerns highlighted by our reviewers. These innovations would support the development of LLMs capable of not only generating accurate scientific responses but also providing researchers with insights into the underlying reasoning, ultimately advancing AI as a robust and trustworthy partner in scientific discovery.

5. Conclusion

This study illuminates both the potential and limitations of current LLMs in addressing intricate scientific inquiries. Claude 3.5 Sonnet stands out for its accuracy and consistency, which suggests that advancements in LLM architecture and specialized training can indeed produce tools that support and, in some cases, increase scientific work. However, our findings also highlight that, across the field, LLMs remain limited by inconsistencies and occasional gaps in domain-specific knowledge.

The significance of this research extends beyond the technical performance of LLMs: it calls for a critical approach to their integration into scientific workflows, demanding transparency, methodological rigor, and ethical foresight. As LLMs become increasingly common in research settings, understanding when and how they should be applied is essential to harness their strengths while mitigating their limitations. The ethical implications of AI in science, highlighted by our reviewers’ feedback, emphasize the need for explicit disclosure and regulation to maintain integrity and trust in AI-assisted findings.

Given the exploratory nature of this work and the limited number of exam questions, the conclusions should be considered preliminary. Larger and more diverse datasets, as well as expanded disciplinary coverage, will be necessary to confirm and extend these findings. Future work should also include replication with independent, blinded evaluators outside the authorship team, to confirm whether the comparative performance patterns observed here remain stable across panels.

Acknowledgments

We thank all the technical and administrative staff who made it possible to carry out the research.

Funding Statement

The author(s) declare that financial support was received for the research and/or publication of this article. This work was supported by RTI2018-096724-B-C22, PID2021-125188OB-C32 and TED2021-129932B-C21 projects funded by Spanish Ministry of Economy and Competitiveness and CB12/03/30038 (CIBER Fisiopatologia de la Obesidad y la Nutrición, CIBERobn, Instituto de Salud Carlos III). This research was also funded by Generalitat Valenciana (grant number PROMETEO/2021/059) and Agencia Valenciana de la Innovación (grant number INNEST/2022/103). FÁ-M were supported by “Margarita Salas” fellowship from the Spanish Ministry of Universities. EB-C was supported by the “Requalification for university teachers grant” from the Spanish Ministry of Universities and European Union Next Generation program.

Data availability statement

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/Supplementary material.

Author contributions

FÁ-M: Conceptualization, Formal analysis, Investigation, Methodology, Software, Validation, Writing – original draft, Writing – review & editing. LE: Investigation, Validation, Writing – original draft, Writing – review & editing. LF: Investigation, Validation, Writing – original draft, Writing – review & editing. EB: Investigation, Validation, Writing – original draft, Writing – review & editing. AZ: Investigation, Validation, Writing – original draft, Writing – review & editing. MH-L: Investigation, Validation, Writing – original draft, Writing – review & editing. MA: Investigation, Validation, Writing – original draft, Writing – review & editing. FP: Investigation, Validation, Writing – original draft, Writing – review & editing. AT: Investigation, Validation, Writing – original draft, Writing – review & editing. JG-P: Investigation, Validation, Writing – original draft, Writing – review & editing. DA-R: Investigation, Validation, Writing – original draft, Writing – review & editing. AR: Investigation, Validation, Writing – original draft, Writing – review & editing. PM: Investigation, Validation, Writing – original draft, Writing – review & editing. RE-E: Investigation, Validation, Writing – original draft, Writing – review & editing. JL: Investigation, Validation, Writing – original draft, Writing – review & editing. VM: Investigation, Validation, Writing – original draft, Writing – review & editing. FB-R: Conceptualization, Data curation, Formal analysis, Methodology, Software, Supervision, Validation, Writing – original draft, Writing – review & editing. EB-C: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Supervision, Validation, Writing – original draft, Writing – review & editing.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The author(s) declared that they were an editorial board member of Frontiers, at the time of submission. This had no impact on the peer review process and the final decision.

Generative AI statement

The authors declare that no Gen AI was used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frai.2025.1664303/full#supplementary-material

Data_Sheet_1.docx (696.3KB, docx)

References

  1. Ali S., Abuhmed T., El-Sappagh S., Muhammad K., Alonso-Moral J. M., Confalonieri R., et al. (2023). Explainable artificial intelligence (XAI): what we know and what is left to attain trustworthy artificial intelligence. Inf. Fusion 99:101805. doi: 10.1016/j.inffus.2023.101805 [DOI] [Google Scholar]
  2. Borger J. G., Ng A. P., Anderton H., Ashdown G. W., Auld M., Blewitt M. E., et al. (2023). Artificial intelligence takes center stage: exploring the capabilities and implications of ChatGPT and other AI-assisted technologies in scientific research and education. Immunol. Cell Biol. 101, 923–935. doi: 10.1111/imcb.12689, PMID: [DOI] [PubMed] [Google Scholar]
  3. Chen T. Challenges and opportunities in integrating LLMs into continuous integration/continuous deployment (CI/CD) pipelines. Proceeding 5th International Seminar Artificial Intelligence, Networking and Information Technology AINIT. Institute of Electrical and Electronics Engineers Inc.; (2024). p. 364–367. [Google Scholar]
  4. Chen D. K., Modi Y., Al-Aswad L. (2022). Promoting transparency and standardization in ophthalmologic artificial intelligence: a call for artificial intelligence model card. Asia Pac. J. Ophthalmol. 11, 215–218. doi: 10.1097/APO.0000000000000469, PMID: [DOI] [PubMed] [Google Scholar]
  5. Cheng S. L., Tsai S. J., Bai Y. M., Ko C. H., Hsu C. W., Yang F. C., et al. (2023). Comparisons of quality, correctness, and similarity between ChatGPT-generated and human-written abstracts for basic research: cross-sectional study. J. Med. Internet Res. 25:e51229. doi: 10.2196/51229, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Deiner M. S., Honcharov V., Li J., Mackey T. K., Porco T. C., Sarkar U. (2024). Large language models can enable inductive thematic analysis of a social media corpus in a single prompt: human validation study. JMIR Infodemiol. 4:e59641. doi: 10.2196/59641, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Elyoseph Z., Levkovich I. (2024). Comparing the perspectives of generative AI, mental health experts, and the general public on schizophrenia recovery: case vignette study. JMIR Ment. Health 11:e53043. doi: 10.2196/53043, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Goel A., Gueta A., Gilon O., Liu C., Erell S., Nguyen L. H., et al. (2023). LLMs accelerate annotation for medical information extraction. Proc. Mach. Learn. Res. 225:296. doi: 10.48550/arXiv.2312.02296 [DOI] [Google Scholar]
  9. Hosseini M., Resnik D. B., Holmes K. (2023). The ethics of disclosing the use of artificial intelligence tools in writing scholarly manuscripts. Res. Ethics 19, 449–465. doi: 10.1177/17470161231180449, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Jahan I., Laskar M. T. R., Peng C., Huang J. X. (2024). A comprehensive evaluation of large language models on benchmark biomedical text processing tasks. Comput. Biol. Med. 171:108189. doi: 10.1016/j.compbiomed.2024.108189, PMID: [DOI] [PubMed] [Google Scholar]
  11. Kim H. S., Kim G. T. (2024). Can a large language model create acceptable dental board-style examination questions? A cross-sectional prospective study. J. Dent. Sci. 20, 895–900. doi: 10.1016/j.jds.2024.08.020 [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Kiseleva A., Kotzinos D., De Hert P. (2022). Transparency of AI in healthcare as a multilayered system of accountabilities: between legal requirements and technical limitations. Front. Artif. Intell. 5:879603. doi: 10.3389/frai.2022.879603, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Kostic M., Witschel H. F., Hinkelmann K., Spahic-Bogdanovic M.. LLMs in automated essay evaluation: a case study. (2024). Available online at: www.aaai.org (Accessed January 13, 2025).
  14. Li L., Li P., Wang K., Zhang L., Ji H., Zhao H. (2024). Benchmarking state-of-the-art large language models for migraine patient education: performance comparison of responses to common queries. J. Med. Internet Res. 26:e55927. doi: 10.2196/55927, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Li X., Ouyang J. A systematic investigation of knowledge retrieval and selection for retrieval augmented generation. Arxiv, Cornell University, New York. (2024). doi: 10.48550/arXiv.2410.13258 [DOI]
  16. Liu K., Chen Z., Fu Z., Zhang W., Jiang R., Zhou F., Chen Y., et al. Structure-aware domain knowledge injection for large language models. Arxiv, Cornell University, New York. (2024). doi: 10.48550/arXiv.2407.16724 [DOI]
  17. Mu X., Lim B., Seth I., Xie Y., Cevik J., Sofiadellis F., et al. (2024). Comparison of large language models in management advice for melanoma: Google’s AI bard, BingAI and ChatGPT. Skin Health Dis. 4:313. doi: 10.1002/ski2.313 [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Perkins M. (2023). Academic integrity considerations of AI large language models in the post-pandemic era: ChatGPT and beyond. J. Univ. Teach. Learn. Pract. 20, 1–24. doi: 10.53761/1.20.02.07 [DOI] [Google Scholar]
  19. Pompili D., Richa Y., Collins P., Richards H., Hennessey D. B. (2024). Using artificial intelligence to generate medical literature for urology patients: a comparison of three different large language models. World. J. Urol. 42:455. doi: 10.1007/s00345-024-05146-3, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Qu Y., Wang J. (2024). Performance and biases of large language models in public opinion simulation. Humanit. Soc. Sci. Commun. 11, 1–13. doi: 10.1057/s41599-024-03609-x [DOI] [Google Scholar]
  21. Rudolph J., Tan S., Tan S. (2023). War of the chatbots: bard, Bing chat, ChatGPT, Ernie and beyond. The new AI gold rush and its impact on higher education. J. Appl. Learn. Teach. 6, 364–389. doi: 10.37074/jalt.2023.6.1.23 [DOI] [Google Scholar]
  22. Ruiz Sarrias O., Del Martínez Prado M. P., Sala Gonzalez M. Á., Azcuna Sagarduy J., Casado Cuesta P., Figaredo Berjano C., et al. (2024). Leveraging large language models for precision monitoring of chemotherapy-induced toxicities: a pilot study with expert comparisons and future directions. Cancers (Basel) 16:2830. doi: 10.3390/cancers16162830, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Saibene A. M., Allevi F., Calvo-Henriquez C., Maniaci A., Mayo-Yáñez M., Paderno A., et al. (2024). Reliability of large language models in managing odontogenic sinusitis clinical scenarios: a preliminary multidisciplinary evaluation. Eur. Arch. Otorrinolaringol. 281, 1835–1841. doi: 10.1007/s00405-023-08372-4, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Sharifi S., Namvar M., Intezari A., Akhlaghpour S. (2024). Healthcare LLMs go to market: a realist review of product launch news. Digit. Health Inform. Innov. Sustain. Health Care Syst. 316, 712–716. doi: 10.3233/SHTI240513 [DOI] [PubMed] [Google Scholar]
  25. Shi R., Liu S., Xu X., Ye Z., Yang J., Le Q., et al. (2024). Benchmarking four large language models’ performance of addressing Chinese patients' inquiries about dry eye disease: a two-phase study. Heliyon 10:e34391. doi: 10.1016/j.heliyon.2024.e34391, PMID: [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Shool S., Adimi S., Saboori Amleshi R., Bitaraf E., Golpira R., Tara M. (2023). A systematic review of large language model (LLM) evaluations in clinical medicine. BMC Med. Inform. Decis. Mak. 25:117. doi: 10.1186/s12911-025-02954-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Wang S., Tan J., Dou Z., Wen J. R.. Omnieval: an omnidirectional and automatic RAG evaluation benchmark in financial domain. Arxiv [Preprint]. (2024). doi: 10.48550/arXiv.2412.13018 [DOI]
  28. Wu X., Saraf P. P., Lee G., Latif E., Liu N., Zhai X. Unveiling scoring processes: dissecting the differences between LLMs and human graders in automatic scoring. Arxiv, Cornell University, New York. (2024). doi: 10.48550/arXiv.2407.18328 [DOI]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Data_Sheet_1.docx (696.3KB, docx)

Data Availability Statement

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/Supplementary material.

Articles from Frontiers in Artificial Intelligence are provided here courtesy of Frontiers Media SA

RESOURCES