How Software Engineering Students Use LLMs to Write Research Papers: An Experience Report Ronnie de Souza Santos1,∗ , Maria Teresa Baldassarre2 , Cleyton Magalhães3 , Italo Santos4 1
University of Calgary, Canada 2 University of Bari, Italy 3 Universidade Federal Rural de Pernambuco (UFRPE), Brazil 4 University of Hawai‘i at Mānoa, USA
arXiv:2606.05114v1 [cs.SE] 3 Jun 2026
[email protected], [email protected], [email protected], [email protected] ∗ corresponding author
Abstract—Large language models are increasingly becoming part of software engineering education, including activities involving empirical software engineering and evidence synthesis. This paper reports an educational experience involving the integration of reflective LLM use into an empirical methods assignment in a third-year software architecture course. Students were asked to develop a short research paper using either a rapid review or a gray literature review methodology and to disclose how LLMs were used throughout the assignment. We analyzed 146 student disclosure statements using a crossanalysis process combining LLM-assisted categorization with manual verification and refinement by the researchers. The reflections describe how students incorporated LLMs during activities such as brainstorming, methodological clarification, organization of findings, and writing refinement, while also reporting concerns regarding inaccuracies and verification of generated content. This experience report discusses lessons learned and educational implications for integrating AI-assisted technologies into empirical software engineering education. Keywords–LLMs; empirical software engineering; undergraduate students 1. I NTRODUCTION Large language models (LLMs) have become increasingly present in educational settings, supporting activities such as tutoring, feedback generation, content creation, and writing assistance [1], [2], [3]. In software engineering education, these tools are now part of students’ everyday academic workflows, including programming activities, brainstorming, information synthesis, and academic writing [4], [5]. At the same time, their growing adoption has raised questions regarding authorship, reliability, critical thinking, and the role of AI in learning processes [6], [7], [8]. Current discussions in software engineering education increasingly argue that LLMs should not be treated only as coding assistants or productivity tools, but also as technologies that require critical and reflective engagement [9], [10]. In this context, educators are beginning to rethink how assignments
are designed, how students are encouraged to interact with AI systems, and how responsible use can be incorporated into teaching practices. These discussions become particularly relevant in empirical software engineering education, where students are expected to develop skills related to evidence synthesis, methodological reasoning, critical analysis, and communication. This paper reports our experience integrating reflective and transparent LLM use into an empirical methods assignment in a third-year software architecture course. In the assignment, students were allowed to use LLMs throughout the process of developing a short research paper, provided that they disclosed how the tools were used. The activity aimed not only to support students during the writing process, but also to encourage reflection about the role of AI tools in academic work. To better understand how students interacted with LLMs in this educational setting, we analyzed 146 disclosure statements written by students after completing the assignment. These statements described how students used LLMs, which aspects of the writing process were supported by the tools, and which benefits and limitations they perceived. Our goal was not to evaluate the effectiveness of LLMs, but rather to reflect on how students appropriated these technologies during an empirical writing activity and what this experience suggests for software engineering education. The observations presented in this paper contribute practical insights for educators interested in incorporating LLMs into empirical software engineering assignments. In particular, our experience highlights the importance of transparency, critical verification, and explicit guidance regarding AI-assisted academic work. From this introduction, this paper is organized as follows. Section 2 discusses empirical methods teaching in software engineering education. Section 3 presents the educational context and the assignment structure. Section 4 reports the main observations from student disclosures. Section 5 discusses lessons learned and implications for education. Finally, Section 7 concludes the paper.
2. T EACHING E MPIRICAL M ETHODS Teaching empirical methods in software engineering is widely recognized as important for helping students understand how to evaluate technologies and practices systematically and evidence-based. The growing role of empirical research in both academia and industry highlights the need to prepare them to assess new methods, tools, and processes before their adoption [11], [12]. These methods, such as controlled experiments, case studies, and surveys, enable students to understand how to make data-driven decisions and analyze and interpret results, reinforcing the role of evaluation in improving software development practices [13]. To address these goals, professors have developed a range of pedagogical strategies. These include embedding empirical assignments into broader software engineering courses, offering specialized empirical methods courses, and using structured project work to simulate evaluation scenarios [11], [14]. Projects often involve students in multiple stages of empirical work, including data analysis, critical reflection on study design, and assessment of threats to validity [11], [13]. Exploring tasks and guiding analysis enables instructors to support student learning across both the theoretical and practical dimensions of empirical research. Nevertheless, implementing such courses involves navigating several challenges. Professors must account for students’ diverse academic backgrounds, which can affect their familiarity with software engineering practices [14]. Moreover, striking a balance between educational objectives and potential research outcomes requires careful planning to ensure that learning remains the primary focus [12]. Despite these constraints, empirical course components have been well received by students and are an important element in software engineering education [11]. 3. E DUCATIONAL C ONTEXT AND E XPERIENCE R EPORT This paper reports our experience incorporating reflective and transparent LLM use into an empirical methods assignment in a third-year undergraduate software architecture course. Rather than evaluating the effectiveness of LLMs themselves, our goal was to better understand how students appropriated these tools while developing academic work and what this experience suggests for software engineering education. 3.1 Educational Context The experience was conducted in a third-year undergraduate software architecture course within a software engineering program. As part of the course activities, students were asked to write a short 2–3-page paper investigating a topic related to software architecture or software design. Students were required to define a research question and apply either a Rapid Review or a Grey Literature Review methodology following structured instructional guidelines provided during the course. For the Rapid Review, students were instructed to systematically identify and analyze at least 15 peer-reviewed studies published in reputable software engineering venues within the last 10 years. In the Grey Literature Review, students were
asked to analyze at least 30 practitioner-oriented online posts from sources such as StackExchange, Quora, and Dev.to. In both activities, students were expected to synthesize findings, organize evidence into categories or themes, and communicate their results using an academic structure. Students were explicitly allowed to use LLMs throughout the assignment development process. This included support for brainstorming, refining research questions, improving writing clarity, organizing findings, and understanding methodological concepts. However, students were instructed that the papers should reflect their own reasoning and synthesis rather than fully AI-generated content. To encourage transparency and reflection, students were required to include a short disclosure statement describing whether and how LLMs were used during the assignment. These disclosures were not graded and were included as part of broader course guidelines regarding AI use across assignments. 3.2 Reflection Collection The disclosure statements were collected only after the course had concluded and final grades had been released. This procedure ensured that no academic consequences could be associated with the disclosures and that students could reflect on their experiences without affecting evaluation outcomes. To preserve anonymity, two Teaching Assistants extracted the disclosure statements from the submitted assignments, removed identifying information, and organized the material into a shared spreadsheet. The statements were inserted in randomized order to reduce the possibility of associating disclosures with specific students. In total, 146 anonymized statements were compiled for reflection and analysis. 3.3 Cross Analysis of Student Reflections To characterize how students described their use of LLMs, we conducted a cross-analysis combining LLM-assisted categorization with manual verification and refinement by the researchers. The process was intentionally designed as a reflective educational analysis rather than a fully automated classification process. The analysis was conducted in four phases. Phase 1: Familiarization with the Reflections. Initially, the researchers independently read the full set of 146 disclosure statements multiple times to become familiar with the reflections and identify recurring patterns related to how students used LLMs during the assignment. During this stage, attention was given to recurring descriptions of writing support, methodological assistance, brainstorming activities, perceived benefits, and reported challenges. Phase 2: Initial LLM-Assisted Categorization. To support the organization of recurring patterns across the disclosures, ChatGPT-4.0 was used to assist with excerpt identification and preliminary categorization tasks. The prompts focused on dimensions commonly emphasized in reviewer guidelines
from major software engineering venues such as ICSE, ICSME, ESEM, EASE, and the Empirical Software Engineering journal. These venues frequently instruct reviewers to evaluate aspects such as novelty, rigor, relevance, transparency, and presentation quality when assessing research papers. Based on these recurring dimensions, the analysis was organized around the following categories: Novelty: uses related to brainstorming, topic exploration, research question definition, and understanding unfamiliar concepts; • Rigor: uses associated with methodological clarification, organization of research procedures, identification of evidence sources, and support for empirical research activities; • Relevance: uses focused on improving the communication, interpretation, coherence, and significance of findings and conclusions; • Transparency: uses related to improving explanations of research procedures, clarity of reporting, traceability of information, and communication of how findings were obtained or analyzed; • Presentation: uses associated with grammar correction, writing refinement, formatting, readability, sentence structure, and overall presentation quality. •
ChatGPT-4.0 was used to identify excerpts potentially associated with these categories and to flag explicit mentions of perceived benefits and challenges related to LLM use. An illustrative example of the prompt structure used during the categorization process is presented below. ¡ Illustrative Analysis Prompt Instruction: Analyze the following student disclosure statements and identify excerpts where students explicitly describe how LLMs supported academic writing activities. Categorize the excerpts according to the following dimensions commonly considered in software engineering peer review: Novelty, Rigor, Relevance, Transparency, and Presentation. Guidelines: Only classify excerpts when the student explicitly describes the use or perceived impact of the tool. Do not infer meanings that are not directly stated in the disclosure statements. Output: Return the identified category together with the corresponding excerpt from the disclosure statement. Phase 3: Manual Verification and Refinement. After the initial AI-assisted categorization, the researchers manually reviewed all classifications and extracted quotations. During this verification process, the researchers checked whether the excerpts accurately reflected the assigned category, removed ambiguous classifications, adjusted labels when necessary, and refined the interpretation of the disclosures to preserve contextual meaning. Particular attention was given to avoiding overinterpretation. Only disclosures in which students explicitly described a use,
challenge, or perception related to LLMs were retained in the final categorization. Statements considered vague, indirect, or unsupported were excluded from the analysis. The researchers also compared classifications collaboratively and discussed disagreements until consensus was reached. Phase 4: Consolidation of Educational Observations. Finally, the refined categories and excerpts were consolidated into broader educational observations regarding how students integrated LLMs into empirical writing activities. At this stage, the researchers focused on recurring educational patterns, including how students perceived LLMs as writing assistants, methodological aids, brainstorming partners, and sources of concern related to hallucinations, inaccuracies, and meaning distortion. This iterative interaction between AI-assisted support and researcher interpretation allowed the organization of recurring educational observations while preserving the contextual richness of the student reflections. 3.4 Ethics All procedures followed institutional ethical guidelines for research involving human participants. Disclosure statements were analyzed only after final grades had been released, ensuring that participation had no influence on academic evaluation. The research team had access only to anonymized disclosures and did not interact with identifiable student data during the analysis process. Due to the nature of the disclosures and the educational setting, the complete dataset cannot be publicly shared. However, anonymized excerpts are included throughout the paper to illustrate the observations discussed. 4. E DUCATIONAL O BSERVATIONS Across the 146 student disclosure statements collected after the assignment, several recurring patterns emerged regarding how students incorporated LLMs into their academic writing activities. Most students described using LLMs to support presentation and writing quality, while smaller groups reported using the tools for idea generation, methodological support, articulation of findings, and transparency-related tasks. A small number of students also explicitly stated that they chose not to use LLMs during the assignment. Most students reported using ChatGPT as their primary LLM during the assignment. A smaller number mentioned other tools, including Claude, Gemini, GitHub Copilot, Bing AI, and Deepseek. In some cases, students described combining multiple LLMs depending on the activity being performed, particularly for brainstorming, writing refinement, or summarization tasks. These observations suggest that while ChatGPT dominated usage patterns, students were also experimenting with different AI tools throughout the assignment process. 4.1 LLMs as Writing and Research Support Tools The most frequently reported use of LLMs involved Presentation. Students commonly described using the tools to revise grammar, improve sentence structure, increase clarity, and
TABLE I I LLUSTRATIVE S TUDENT D ISCLOSURES BY T YPE OF LLM U SE Presentation “This paper leveraged ChatGPT and Google Gemini for the improvement of sentence structure, spelling, grammar, and Latex formatting.” (D001) “ChatGPT, a generative AI tool, was used in the writing process of this paper for grammatical checks and proper sentence phrasing.” (D005) “In this paper, ChatGPT and Grammarly AI were tools used for grammar, spelling/thesaurus, checking if content was in accordance with IEEE standard.” (D006) Novelty “In the early brainstorming stages of this paper, ChatGPT was used to gain a broad understanding (...)” (D013) “The author of this paper used ChatGPT-4o to generate ideas for the topic of this paper.” (D053) “I used Chatgpt to understand the key concepts of the papers.” (D120) Relevance “(...) it helped me word and phrase my findings and thoughts in a more cohesive and clear manner.” (D016) “It helped me break down complicated ideas, explain difficult terms, and ensure that my explanations made sense..” (D111) “In this paper, ChatGPT was used to (...) refining clarity in written explanations” (D134) Rigor “(...) to create a step-by-step guide for conducting research using the prompt “can you devise a plan to research and write the paper based on that question?’” (D024) “In this paper, the DeepSeekR1 LLM served as a research partner in this paper, providing a second perspective during the study screening process by evaluating inclusion criteria that I originally created.” (D099) “Furthermore, I used it to give me key bullet points from some of the papers I researched.” (D133) Transparency “A blueprint of the structure of this paper was provided, to which the tool identified areas where additional material could enhance the explanation..” (D083) “In this paper, ChatGPT was used to enhance the reader’s comprehension (...) making the paper more transparent and effective.” (D093)
questions, and better understand unfamiliar concepts related to software architecture and software design. Some students reported iterating over multiple prompts to compare possible directions for their papers or to better define the scope of their investigations. In several disclosures, students described the tools as interactive brainstorming partners that helped initiate early-stage academic reasoning and topic exploration. A smaller group of students described using LLMs to support Relevance. In these cases, students reported asking the tools for assistance in explaining the relevance of their observations, synthesizing conclusions, or improving the communication of implications identified during the assignment. Rather than generating findings directly, the tools were often used to help students organize and phrase their interpretations more clearly. These reflections suggest that some students perceived LLMs as support mechanisms for strengthening coherence and improving how results and implications were communicated in their papers. Some students also described using LLMs to support Rigor. These uses included asking for clarification about rapid reviews and gray literature reviews, obtaining suggestions for structuring research procedures, and improving descriptions of data organization and analysis activities. A few students also reported using the tools to help identify possible evidence sources and organize categories or themes across collected materials. From an educational perspective, these observations indicate that students may perceive LLMs not only as writing assistants, but also as tools capable of supporting understanding of empirical software engineering practices and research organization activities. Finally, a small number of students reported using LLMs to support Transparency. In these cases, students described using the tools to improve explanations regarding how information was collected, categorized, or analyzed. Some students also used LLMs to review whether their descriptions of research procedures were sufficiently understandable and reproducible. Although less frequent, these observations suggest that some students were already reflecting on how AI tools could support communication of research processes, methodological clarity, and transparency during academic writing activities. 4.2 Student Reflections on Benefits and Challenges
adapt the tone of their papers to a more academic style. Several students characterized the LLM as a “final editor” used before submission to polish the text and improve readability. Students also reported using LLMs to support citation formatting, particularly in BibTeX and LATEX, and to help reorganize sections to better align with the assignment template. These observations suggest that students frequently relied on LLMs to navigate academic writing conventions and improve the overall presentation of their work. Another recurring pattern involved the use of LLMs for Novelty. Students described using the tools to explore possible research topics, narrow broad themes, identify potential research
Students reported a wide range of experiences regarding the incorporation of LLMs into academic writing activities. While some disclosures were brief and technical, others provided detailed reflections about how these tools affected writing organization, comprehension, confidence, and learning processes throughout the assignment. Many students described LLMs as useful companions during the writing process. Commonly reported benefits included improved grammar, better organization of ideas, assistance with summarization, and support for understanding complex concepts. Several students also emphasized that LLMs helped make academic writing feel more manageable, particularly when dealing with unfamiliar
terminology, empirical methods, or formal academic structures. For example, student D138 stated: “AI did help improve my writing process by helping me to avoid repeating the same ideas again and again.” Similarly, student D109 explained: “Additionally, it provided support in understanding complex concepts and rephrasing them in simpler terms so I could grasp them better.” These reflections suggest that many students perceived LLMs as tools capable of supporting both communication and comprehension during empirical writing activities. In particular, several disclosures associated AI support with increased clarity, confidence, organization, and reduced difficulty during the writing process. At the same time, students also expressed concerns regarding limitations and risks associated with LLM use. Several disclosures mentioned hallucinations, inaccuracies, lack of contextual understanding, and the possibility that AI-generated suggestions could unintentionally distort the intended meaning of the text. Some students also described difficulties determining whether generated outputs were reliable or sufficiently aligned with their intended arguments. These concerns indicate that, despite the perceived usefulness of the tools, students remained aware of practical limitations and risks associated with relying on AI-generated content during academic work. For example, participant D026 stated: “A key challenge being that the AI-assisted refinements altered the intended meaning or introduce unintended biases.” Another student, D018, reported: “However, hallucinations were an issue as in certain cases it made up fake information in terms of certain topics.” These disclosures suggest that students were not simply accepting generated outputs without reflection. Instead, many students described the need to verify information, critically evaluate generated suggestions, and preserve authorship and meaning during the writing process. Considering the reflections as a whole, the statements revealed a combination of perceived benefits and challenges associated with the use of LLMs in academic writing activities. 5. L ESSONS L EARNED AND E DUCATIONAL I MPLICATIONS Empirical software engineering plays an important role in software engineering education because students are increasingly expected to synthesize evidence, evaluate technologies, and communicate empirical findings using structured methodologies [11], [12], [13]. One lesson learned from this experience was that many students perceived LLMs as writing companions rather than simple correction tools. Although presentation improvement was the most frequently reported use, students also described using LLMs to brainstorm ideas, organize arguments, clarify concepts, structure findings, and refine explanations throughout the writing process. Several disclosures characterized the tools as interactive partners that helped students navigate empirical writing activities and better understand academic expectations. Another lesson learned involved how students balanced the perceived usefulness of LLMs with concerns regarding reliability and authorship. Many disclosures mentioned hallu-
cinations, inaccuracies, unintended changes in meaning, and difficulties evaluating generated outputs. At the same time, students described reviewing, revising, and verifying generated suggestions before incorporating them into their papers. These observations suggest that students were not simply outsourcing academic work to AI systems, but instead using LLMs as supportive writing partners while attempting to maintain human judgment and control over the final content. Finally, this experience suggests that reflective disclosure activities may support more transparent and critical engagement with AI-assisted technologies in software engineering education. Asking students to explicitly describe how LLMs supported their work encouraged reflections not only about benefits, but also about limitations, confidence, uncertainty, and responsibility during empirical writing activities. 6. L IMITATIONS The main limitation of this paper is that it reports observations from a single undergraduate software engineering course and one specific empirical writing assignment. Consequently, the observations may not reflect how students use LLMs in other courses, institutions, or educational contexts. A secondary limitation is that the disclosures represent self-reported reflections written after the assignment was completed. Some statements were detailed, while others were brief, which may have influenced the depth of the analysis. Another limitation is that the study focused on disclosure statements rather than direct observation of how students interacted with LLMs during the assignment process. Additional data sources, such as interviews or classroom observations, could provide a broader understanding of how students incorporate AI tools into empirical software engineering activities. 7. C ONCLUSIONS This paper reports an educational experience involving the integration of reflective LLM use into an empirical software engineering assignment in an undergraduate software engineering course. Through the analysis of 146 student disclosure statements, we explored how students described the incorporation of LLMs while conducting activities related to rapid reviews and gray literature reviews. The analysis focused on five recurring dimensions of use: Presentation, Novelty, Relevance, Rigor, and Transparency. The reflections showed that students used LLMs not only for writing refinement activities, but also for brainstorming, methodological clarification, organization of findings, and communication of empirical results. At the same time, students frequently reported concerns involving hallucinations, inaccuracies, meaning distortion, and the need to verify generated outputs. These observations suggest that students are already incorporating AI tools into empirical software engineering workflows while negotiating issues related to authorship, reliability, and critical evaluation. For future work, we plan to conduct additional empirical studies involving different software engineering courses, instructional strategies, and evidence synthesis activities, such as
systematic literature reviews and multivocal literature reviews, to better understand how students incorporate AI-assisted technologies into empirical software engineering workflows and learning processes. R EFERENCES [1] S. Moore, R. Tong, A. Singh, Z. Liu, X. Hu, Y. Lu, J. Liang, C. Cao, H. Khosravi, P. Denny et al., “Empowering education with llms-the next-gen interface and content generation,” in International Conference on Artificial Intelligence in Education. Springer, 2023, pp. 32–37. [2] Q. Wen, J. Liang, C. Sierra, R. Luckin, R. Tong, Z. Liu, P. Cui, and J. Tang, “Ai for education (ai4edu): Advancing personalized education with llm and adaptive learning,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 6743–6744. [3] S. Wiktor, M. Dorodchi, and N. Wiktor, “Ai can help instructors help students: An llm-supported approach to generating customized student reflection responses,” in 2024 IEEE Frontiers in Education Conference (FIE). IEEE, 2024, pp. 1–9. [4] S. Filippi and B. Motyl, “Large language models (llms) in engineering education: A systematic review and suggestions for practical adoption,” Information, vol. 15, no. 6, p. 345, 2024. [5] U. Arora, A. Garg, A. Gupta, S. Jain, R. Mehta, R. Oberoi, Prachi, A. Raina, M. Saini, S. Sharma et al., “Analyzing llm usage in an advanced computing class in india,” in Proceedings of the 27th Australasian Computing Education Conference, 2025, pp. 154–163. [6] L. Yan, L. Sha, L. Zhao, Y. Li, R. Martinez-Maldonado, G. Chen, X. Li, Y. Jin, and D. Gašević, “Practical and ethical challenges of large language models in education: A systematic scoping review,” British Journal of Educational Technology, vol. 55, no. 1, pp. 90–112, 2024.
[7] V. Grande, N. Kiesler, and M. A. Francisco R, “Student perspectives on using a large language model (llm) for an assignment on professional ethics,” in Proceedings of the 2024 on Innovation and Technology in Computer Science Education V. 1, 2024, pp. 478–484. [8] M. A. Razafinirina, W. G. Dimbisoa, and T. Mahatody, “Pedagogical alignment of large language models (llm) for personalized learning: a survey, trends and challenges,” Journal of Intelligent Learning Systems and Applications, vol. 16, no. 4, pp. 448–480, 2024. [9] V. D. Kirova, C. S. Ku, J. R. Laracy, and T. J. Marlowe, “Software engineering education must adapt and evolve for an llm environment,” in Proceedings of the 55th ACM Technical Symposium on Computer Science Education V. 1, 2024, pp. 666–672. [10] J. Pereira, J.-M. López, X. Garmendia, and M. Azanza, “Leveraging open source llms for software engineering education and training,” in 2024 36th International Conference on Software Engineering Education and Training (CSEE&T). IEEE, 2024, pp. 1–10. [11] M. Host, “Introducing empirical software engineering methods in education,” in Proceedings 15th Conference on Software Engineering Education and Training (CSEE&T 2002). IEEE, 2002, pp. 170–179. [12] C. Wohlin, “Empirical software engineering: Teaching methods and conducting studies,” in Empirical Software Engineering Issues. Critical Assessment and Future Directions: International Workshop, Dagstuhl Castle, Germany, June 26-30, 2006. Revised Papers. Springer, 2007, pp. 135–142. [13] S. Pizard, F. Acerenza, X. Otegui, S. Moreno, D. Vallespir, and B. Kitchenham, “Training students in evidencebased software engineering and systematic reviews: a systematic review and empirical study,” Empirical Software Engineering, vol. 26, pp. 1–53, 2021. [14] A. Serebrenik and N. Cassee, “Teaching empirical methods at eindhoven university of technology,” in Handbook on Teaching Empirical Software Engineering. Springer, 2024, pp. 179–207.