US China AI research collaboration: Small Share, Outsized Impact at the Frontier

By Agustin Giovagnoli / January 21, 2026

The data points to a paradox: policy rhetoric leans toward decoupling, but practical research ties between the US and China continue to knit together around core AI architectures, benchmarks, and open-source models. For leaders weighing US China AI research collaboration, the signal is that even a small share of coauthored work can shape frontier directions—and business roadmaps—disproportionately [1][2][3].

Quick summary of US China AI research collaboration

At NeurIPS, one of the field’s premier venues, about 3% of papers include coauthors from both US and Chinese institutions—141 of 5,290 in a recent year, and 134 of 4,497 in 2024—an enduring, if narrow, channel for cross-border work [1]. These collaborations frequently converge on shared use of transformer architectures and open-source LLMs such as Meta’s Llama, as well as Chinese-developed models like Alibaba’s Qwen, reflecting practical, interoperable toolchains for cross-Pacific teams [1]. Beyond conferences, large-scale bibliometrics show the US and China jointly dominate global productivity, novelty, and impact, with cross-country papers often delivering higher research influence [2].

The data: What conferences and bibliometrics show

• NeurIPS US China coauthorship: ~3% of accepted papers in recent years (141/5,290; 134/4,497) [1].
• Team structure: US–China collaborative AI papers generally involve more authors than single-country work [1].
• Institutional strength: China-based papers with US collaborators more often feature last authors affiliated with top-100 AI institutions [1].
• Impact: Joint papers are associated with higher research impact based on large-scale analyses [2].
• Chinese NeurIPS growth: Accepted papers from Chinese entities grew 3.77x from 2021 to 2024, with more than 3,700 accepted over four years, and several China-based labs now among top contributors [3].

For context on the venue’s influence, see the official NeurIPS website (external).

What technologies and models connect teams

The connective tissue is technical: transformer research US China overlaps heavily, and cross-continental teams often standardize on open-source LLM collaboration US China workflows using Meta’s Llama and Alibaba’s Qwen [1]. These common architectures and model families underpin shared experiments, datasets, and benchmarks, enabling reproducibility and faster iteration across institutions [1].

Mechanisms keeping ties open

Migration patterns matter. Scientists moving between the US and China frequently preserve coauthorship networks, which helps maintain collaboration channels and contributes to cross-border AI research impact despite geopolitical headwinds [2]. That continuity, combined with the field’s open-source norms, provides resilience even as policies tighten [2].

Case examples at NeurIPS

Recorded Future’s assessment highlights China’s rapid rise in NeurIPS acceptances and the prominence of several Chinese labs among top contributors [3]. Some award-winning work at the conference involved China–US or China–US-linked teams, including collaborations connected to Microsoft Research Asia and Chinese universities and labs [3]. This momentum underscores how outcomes at elite venues can be shaped by cross-border partnerships even when overall coauthorship remains a small fraction [3].

Strategic implications for businesses and product teams

• Model sourcing and interoperability: Shared reliance on transformers, Llama, and Qwen suggests practical options for evaluation pipelines, fine-tuning, and benchmarking across teams or vendors connected to both ecosystems [1].
• Talent strategy: Larger team sizes and higher-impact outputs in cross-border collaborations signal the value of recruiting from or partnering with groups experienced in multinational research [1][2].
• Portfolio bets: The evidence that joint papers tend to be more impactful supports selectively engaging with cross-border institutions where governance allows [2].
• Monitoring Chinese NeurIPS growth: Rising acceptance volumes and top-lab contributions indicate expanding capability; product roadmaps should track methods and datasets gaining traction in these communities [3].

For hands-on guidance in turning research into products, Explore AI tools and playbooks.

Policy and risk considerations

While debates over decoupling persist, practical research remains materially interconnected through shared architectures, benchmarks, and open-source ecosystems [1][3]. Firms should monitor evolving export controls, collaboration policies, and institutional affiliations to assess exposure and execution risk—especially when partnering with teams straddling jurisdictions [1][3].

Takeaways and next steps

• The slice is small but strategic: US–China coauthorship at NeurIPS hovers around 3%, yet centers on frontier methods and models with disproportionate influence [1].
• Impact edge: Bibliometrics indicate that cross-border AI research impact tends to be higher, reinforcing the value of carefully governed collaborations [2].
• Track acceleration: Chinese NeurIPS growth and award recognition signal continued capability expansion with direct relevance to product and research planning [3].
• Action list: Audit model and data dependencies for cross-border ties; watch conference affiliations and award trends; and build policies to evaluate opportunities in US China AI research collaboration without overexposure to regulatory shifts [1][2][3].

Sources

[1] The US and China Are Collaborating More Closely on AI Than You …
https://www.wired.com/story/us-china-collaboration-neurips-papers/

[2] China and the U.S. produce more impactful AI research … – Nature
https://www.nature.com/articles/s41598-024-79863-5

[3] [PDF] Measuring the US-China AI Gap – Recorded Future
https://assets.recordedfuture.com/insikt-report-pdfs/2025/ta-2025-0508.pdf