How Software Engineering Students Use LLMs to Write Research Papers: An Experience Report

Technical Summary: How Software Engineering Students Use LLMs to Write Research Papers

Problem Statement
The rapid integration of Large Language Models (LLMs) into software engineering education has raised critical questions regarding authorship, reliability, critical thinking, and the pedagogical role of AI in learning processes. While LLMs are increasingly used for coding and writing assistance, there is a lack of empirical understanding regarding how undergraduate students appropriate these tools specifically within empirical software engineering tasks. These tasks require evidence synthesis, methodological reasoning, and critical analysis, areas where the balance between AI assistance and student agency is particularly complex. Educators need to understand how students interact with LLMs during empirical writing to design assignments that foster responsible use and critical engagement rather than mere productivity.

Methodology
This paper presents an experience report based on a third-year undergraduate software architecture course. The study focused on an assignment where students were required to write a short research paper (2–3 pages) using either a Rapid Review or a Gray Literature Review methodology.

• Context: Students were explicitly permitted to use LLMs throughout the process (brainstorming, methodological clarification, organization, writing refinement) provided they maintained transparency.
• Data Collection: Students were required to submit a disclosure statement detailing how LLMs were used. These statements were collected only after final grades were released to eliminate academic bias.
• Sample: The analysis included 146 anonymized disclosure statements.
• Analysis Process: The researchers employed a four-phase cross-analysis approach:
  1. Familiarization: Researchers manually read all statements to identify patterns.
  2. LLM-Assisted Categorization: ChatGPT-4.0 was used to identify excerpts and perform preliminary categorization based on five dimensions derived from major software engineering venue guidelines (Novelty, Rigor, Relevance, Transparency, Presentation).
  3. Manual Verification: Researchers manually reviewed, refined, and verified all AI-generated classifications to ensure accuracy and avoid overinterpretation.
  4. Consolidation: Refined categories were synthesized into educational observations regarding student appropriation of the technology.

Key Contributions and Results
The analysis of the 146 disclosures revealed that students utilized LLMs across five distinct dimensions, moving beyond simple grammar correction:

1. Presentation (Most Frequent): The majority of students used LLMs for grammar correction, sentence structure improvement, tone adaptation to academic styles, and formatting (e.g., BibTeX, LaTeX). Many viewed the LLM as a "final editor."
2. Novelty: Students used tools for brainstorming topics, defining research questions, and understanding unfamiliar concepts in software architecture.
3. Rigor: A subset of students used LLMs to clarify methodological concepts (e.g., rapid vs. gray literature reviews), structure research procedures, and organize evidence sources.
4. Relevance: Students utilized tools to help phrase findings, synthesize conclusions, and improve the coherence of their arguments.
5. Transparency: A smaller group used LLMs to review the clarity and reproducibility of their research procedure descriptions.

Student Perceptions of Benefits and Challenges:

• Benefits: Students reported that LLMs made academic writing more manageable, aided in understanding complex terminology, improved organization, and reduced the difficulty of navigating formal academic structures.
• Challenges: Significant concerns were raised regarding hallucinations, factual inaccuracies, and the risk of AI-generated content distorting the intended meaning or introducing unintended biases.
• Behavior: The data suggests students did not simply outsource work; rather, they acted as critical verifiers, reviewing and revising AI outputs to maintain authorship and accuracy.

Significance and Claims
The paper modestly claims to offer practical insights for educators integrating AI into empirical software engineering assignments, rather than definitive proof of LLM efficacy. The primary significance lies in demonstrating that:

• Transparency is feasible: Requiring disclosure statements encourages students to reflect critically on their AI use without compromising their grades.
• Role Expansion: Students perceive LLMs not just as writing assistants but as interactive partners for methodological reasoning and idea generation.
• Educational Implication: The experience suggests that explicit guidance on transparency, critical verification, and the limitations of AI is essential. It highlights that students are already negotiating issues of authorship and reliability in their workflows, necessitating a shift in pedagogy to support critical engagement with AI tools in empirical research.

The authors acknowledge limitations, noting that the findings are specific to a single course and rely on self-reported data rather than direct observation of the writing process. Future work is proposed to explore these dynamics across different courses and evidence synthesis activities.