Shallow Synthesis of Knowledge in GPT-Generated Texts: A Case Study in Automatic Related Work Composition

1 Introduction

• Scientific literature expanded at 5.08% annually from 1952 to 2020, doubling every 14 years; machine-learning and AI papers nearly doubled every two years from 1994 to 2022.The paper presents this growth as context for the development of AI-assisted scholarly applications.Source: §1 · p. 1
• AI-assisted applications support literature discovery, reading, and writing, and scholars also use ChatGPT directly in academic writing.Savelka et al. (2023) found 1,226 publications acknowledging ChatGPT use between 2022 and 2023.Source: §1 · p. 1
• The literature-review standard used here emphasizes connecting, synthesizing, and sometimes evaluating studies while framing the new work’s relevance and research gap.Quality also depends on which works authors choose to discuss together and how those choices support the framing.Source: §1 · pp. 1–2
• A one-to-one comparison of citation occurrences is inadequate because valid related-work sections depend on the paper’s context and author goals.Classifying citation sentences as true or false positives or negatives is difficult to automate and may involve subjectivity.Source: §1 · p. 2
• Citation graphs measure how often cited works are connected within the same sentence, offering a reproducible, scalable proxy for synthesis and contextual integration.The method represents cited works as nodes and co-occurrences in a sentence as edges.Source: §1 · p. 2
• ScholaCite organizes citations and helps compose related-work sections, while the paper’s contributions include a citation-graph evaluation method and insights into LLM capabilities and limitations.The authors aim to assess whether AI text can perform the synthesis and contextual framing expected in quality academic writing.Source: §1 · pp. 1–2

Automatic Related Work Generation Approaches

• Automatic related-work generators use approaches including sentence extraction and reordering, document relations, contrastive learning, topic modeling, and citation-content planning.Examples include SERGE, ToC-RWG, and systems that structure citations before lexicalization.Source: §2 · p. 2

Automated Systematic Literature Review Applications

• Prior systematic-review tools extract and summarize information, integrate AI methods, or use knowledge graphs to represent literature relationships.This paper instead evaluates outputs of an AI-assisted related-work generation system using citation-graph analysis.Source: §2 · p. 3

The Use of ChatGPT in Academia

• Discussion of ChatGPT in academia balances potential gains in creativity and productivity against accuracy, plagiarism, and academic-integrity concerns.This paper contributes by analyzing GPT-4’s ability to synthesize cited works in scholarly writing.Source: §2 · p. 3

3.1 Task

• Ten works-in-progress were evaluated under original human-written, human-interactive ScholaCite, and GPT-4-only conditions.Each work-in-progress paired a paper with abstracts of works cited in its original related-work section; the original section was redacted for generation.Source: §3.1 · pp. 3–4

3.2 ScholaCite Application

• ScholaCite offers GPT-4 citation-group suggestions and rationales, then lets authors refine those groups before text generation.Grouping and drafting form a two-step workflow; GPT-4 was used to pre-generate groupings for this study, and authors could edit them and the resulting text.Source: §3.2 · p. 4

Step 2: Generation of Related Work Section

• GPT-4 drafts a distinct paragraph for each completed citation group, using its name, citations, and rationale to synthesize related works and frame the work-in-progress.The groupings are supplied in JSON format.Source: Step 2: Generation of Related Work… · p. 5

3.3 Data

• The dataset comprises ten award-winning 2023 ACL papers in English and NLP, selected to postdate GPT-4’s training cutoff; original related-work sections contained 8–52 citations.Source: §3.3 · p. 5

3.3.1 Preprocessing

• Preprocessing represented each work-in-progress with its title, abstract, introduction, and conclusion, and limited the citation list to works cited in its related-work section.Citation records included bibliographic metadata such as titles, abstracts, authors, and years.Source: §3.3.1 · p. 5

3.4 Procedures

• For assisted samples, an author reviewed the work-in-progress and cited abstracts, refined ScholaCite’s groupings, and edited its draft; GPT-only samples used the same sequence without human intervention.Source: §3.4 · pp. 5–6

3.5 Evaluation: Citation Graph Analysis

• Citation graphs treated citations as nodes and same-sentence co-occurrences as edges, then compared edge count, average degree, density, and clustering coefficient across texts.The study used Mann–Whitney U tests for differences between independent samples and applied a Holm–Bonferroni correction for multiple comparisons.Source: §3.5 · p. 6

4 Results

• 48.3 edges in original texts compared with 12.3 for GPT-4 and 38.5 for GPT-assisted sections.Source: §4 · p. 6
• 3.23 average node degree in human-written texts compared with 0.7 for GPT-generated and 2.69 for GPT-assisted texts.The accompanying interpretation says citations in human-written and GPT-assisted texts more often connected to multiple works rather than appearing in isolation.Source: §4 · pp. 6–7
• In the example about PLM reasoning, the original text compared reasoning types in greater detail, while GPT-generated and assisted samples showed less detailed comparison.The GPT-generated sample still included some synthesis; the assisted sample had a similar level of synthesis to GPT-4 for this paper.Source: Table 3 · p. 7
• 0.14 density in original texts and 0.13 in GPT-assisted texts, compared with 0.02 in GPT-generated texts.All texts had sparse citation networks overall.Source: §4 · p. 7
• 0.66 average clustering coefficient in original texts and 0.63 in GPT-assisted texts, compared with 0.15 for GPT.The authors interpret this as greater clustering of citations into subgroups in human-written and collaborative sections.Source: §4 · p. 7
• Across edge count, node degree, density, and clustering, human-written sections showed more interconnected citation structures than GPT-generated sections, with assisted sections often closer to human results.Source: §4 · pp. 6–7

5 Conclusion

• With appropriate, detailed prompting and scholar oversight, LLMs may play a beneficial assisting role.The authors suggest that fine-tuning specifically for synthesis tasks could improve performance.Source: §5 · p. 8
• Without human intervention, GPT-4 fails to interconnect citations to the degree seen in human-written sections, while ScholaCite’s collaborative approach yields citation networks indistinguishable from human expert writing.The authors suggest this reflects limits on deep scholarly knowledge integration from abstract-level data; they also note that context-window limits restrict connections across very long texts.Source: §5 · p. 8

6 Ethical Considerations

• Fully automated scholarly writing risks compromising knowledge synthesis and citation integration, while LLM use also raises concerns about propagated inaccuracies and plagiarism.The authors identify transparency, copyright protections, verification, and semantically meaningful prompting as relevant safeguards.Source: §6 · p. 8

7 Limitations

• The sample of 10 papers, all best papers from the 2023 ACL proceedings, limits generalizability and does not represent the full diversity of NLP.The authors recommend follow-up studies across other fields and venue types.Source: §7 · p. 8
• Citation-graph analysis is a scalable proxy for cited-work synthesis, not a measure of every facet of quality or the nuance of scholarly discourse.The authors note that even highly interconnected citation graphs may lack deeper conceptual integration, and that qualitative assessment could reveal additional capabilities and deficiencies.Source: §7 · pp. 8–9
• The evaluation examines only one set of hyperparameters and prompt-design choices; varying context-window size and fine-tuning approaches could potentially improve synthesis.Source: §7 · p. 9
• The paper does not evaluate ScholaCite itself; only one author used the tool, and a human study is reserved for future work.Source: §7 · p. 9

A.1.1 Step 1a

• Step 1a asks GPT-4 to organize works into thematic groups using the in-progress paper’s abstract, introduction, conclusion, and related-work list.Source: §A.1.1 · p. 11
• Each Step 1a group receives a descriptive name, a rationale for its framing role, and a list of cited papers; every article must appear at least once but may appear in multiple groups.Source: §A.1.1 · p. 11

A.1.2 Step 1b

• Step 1b takes the work-in-progress, existing groups, and cited-work abstracts, then asks for rationales and supporting abstract spans for each cited work.Source: §A.1.2 · pp. 11–12
• The Step 1b output retains the group structure and adds citation rationales and the supporting abstract span to each cited work.Source: §A.1.2 · p. 12

A.2 Prompt for Step 2: Generation of Related Work Section

• Given the context of an in-progress scholarly paper and groupings of related works provided in JSON format, the task is to generate the work-in-progress paper’s related-work section.Source: §A.2 · p. 13
• The generation prompt represents each group with a name, rationale, and cited-paper list, with each paper carrying an ID and title.Source: §A.2 · p. 13
• For Step 2, each cited-paper entry includes its citation ID, title, rationale for group inclusion, and the supporting abstract span.Source: §A.2 · p. 13
• Step 2 requires each grouping to become a distinct paragraph and instructs the model to cite papers by citation ID without referring to the group structure.Source: §A.2 · p. 13

• The graphs are labeled Human-Generated, AI-Assisted, and GPT-Generated.Source: B Graph Representations · pp. 14–15
• Table 5 presents citation graphs for samples 1–5.Source: Table 5 · p. 14
• Table 6 presents citation graphs for samples 6–10.Source: Table 6 · p. 15