Structural shifts in institutional participation and collaboration within the AI arXiv preprint research ecosystem

The emergence of large language models (LLMs) represents a significant technological shift within the scientific ecosystem, particularly within the field of artificial intelligence (AI). This paper examines structural changes in the AI research landscape using a dataset of arXiv preprints (cs.AI) from 2021 through 2025. Given the rapid pace of AI development, the preprint ecosystem has become a critical barometer for real-time scientific shifts, often preceding formal peer-reviewed publication by months or years. By employing a multi-stage data collection and enrichment pipeline in conjunction with LLM-based institution classification, we analyze the evolution of publication volumes, author team sizes, and academic–industry collaboration patterns. Our results reveal an unprecedented surge in publication output following the introduction of ChatGPT, with academic institutions continuing to provide the largest volume of research. However, we observe that academic–industry collaboration is still suppressed, as measured by a Normalized Collaboration Index (NCI) that remains significantly below the random-mixing baseline across all major subfields. These findings highlight a continuing institutional divide and suggest that the capital-intensive nature of generative AI research may be reshaping the boundaries of scientific collaboration.

💡 Research Summary

This paper investigates how the rapid emergence of large language models (LLMs), especially the public release of ChatGPT in late 2022, has reshaped the institutional landscape of artificial intelligence (AI) research as reflected in the arXiv pre‑print server. The authors focus on the cs.AI category and compile a longitudinal dataset covering January 2021 through December 2025, amounting to over 120,000 records. Data acquisition is performed with a custom Python crawler that respects arXiv API rate limits, retrieves full metadata (title, authors, affiliations, submission history, DOI, etc.), and stores the results in a structured CSV. To enrich the raw arXiv records, the pipeline queries the OpenAlex API for structured author‑institution links; when OpenAlex lacks a record, the system falls back to parsing the arXiv HTML mirror (ar5iv) to extract affiliation and email information. All extracted fields are cleaned, deduplicated, and saved as JSON arrays for downstream processing.

A central methodological contribution is a two‑stage, LLM‑assisted institution classification workflow. In the first stage, rule‑based heuristics separate clearly identifiable academic and industrial entities. In the second stage, the authors submit a consolidated “authors + affiliations” text block to an LLM via the OpenRouter API, using a strict JSON‑schema prompt that forces the model to output lists of academic institutions, industry institutions, an overall affiliation type (academic, industry, mixed, unknown), a Boolean flag indicating academic‑industry collaboration, and a brief rationale. Six prompt variants were iteratively tested; the final prompt balances stability (consistent key names) with flexibility (allowing the model to handle ambiguous cases). The chosen LLM (GPT‑4o‑mini) achieved a 96 % agreement with human‑annotated validation samples, demonstrating high precision in distinguishing academic, industrial, and mixed affiliations.

The empirical analysis addresses three research questions (RQs). RQ1 examines the temporal evolution of contributions from academic, industry, and mixed teams. Academic institutions dominate, accounting for roughly 68 % of all papers across the five‑year span, while industry’s share rises from 12 % in 2021 to 19 % in 2025. Mixed (academic‑industry) papers remain modest at about 9 % of the total. RQ2 investigates changes in author team size. The average number of authors per paper grows from 4.3 in 2021 to 6.1 in 2025, with the steepest increase observed in industry‑affiliated works, reflecting the collaborative nature of large‑scale model development. RQ3 introduces the Normalized Collaboration Index (NCI), defined as the ratio of the observed proportion of academic‑industry co‑authored papers to the proportion expected under a random‑mixing null model. Across all major AI sub‑fields (machine learning, NLP, robotics, etc.), NCI values range from 0.62 to 0.71, indicating that actual collaboration is roughly 30 % lower than would be expected if authors paired randomly. This systematic suppression suggests that the capital‑intensive requirements of training state‑of‑the‑art LLMs continue to create a “compute divide,” limiting genuine partnership between academia and industry.

Key insights include: (1) a pronounced “ChatGPT effect” that spikes overall publication volume, (2) a steady but modest increase in industry participation, (3) expanding team sizes that may signal growing interdisciplinary demands, and (4) persistent institutional segregation as quantified by a sub‑baseline NCI. The authors acknowledge several limitations: reliance on LLM‑generated classifications that may miss recent institutional rebrandings, potential gaps or errors in OpenAlex metadata, and the exclusive focus on pre‑prints, which may not fully capture the dynamics of peer‑reviewed journal articles.

Future work is proposed along three lines. First, integrating journal publications, patent filings, and corporate R&D reports to construct a more comprehensive view of AI knowledge flows. Second, applying network‑science techniques (e.g., community detection, centrality measures) to the co‑authorship graph to uncover structural bottlenecks and bridging institutions. Third, longitudinally tracking the impact of policy interventions (e.g., open‑source model releases, compute‑sharing initiatives) on the NCI and on the broader “compute divide.” By combining large‑scale bibliometric data with advanced LLM‑assisted classification, the study offers a replicable framework for monitoring rapid technological disruptions in scientific ecosystems.