Should Chinese Researchers Still Submit to NeurIPS? A Decision‑Utility Analysis

Background

In March 2026 the NeurIPS 2026 submission guidelines added a clause requiring compliance with U.S. law, effectively prohibiting researchers from institutions listed on OFAC sanctions (e.g., Huawei, SenseTime, Harbin Institute of Technology, Beihang University) from submitting, reviewing, or holding positions at the conference. The Chinese Computer Federation (CCF) and the China Science & Technology Association (CSTA) responded by urging a boycott and stating that NeurIPS papers would no longer count as "representative achievements" for domestic projects.

NeurIPS later issued an apology, calling the inclusion of the sanctions link a "mistake" and asserting that the submission rules remained unchanged. The key question is whether this apology alters the incentive structure for Chinese researchers.

Individual Decision Model

Researchers choose between two options: submit to NeurIPS or to alternative top conferences (e.g., ICLR, ICML). The expected utility of submitting to a venue can be expressed as:

Acceptance probability – roughly 25 % for NeurIPS.

Academic reputation – the value of peer recognition and citation potential.

Career value – the tangible benefit for promotions, grant applications, and other institutional evaluations.

Weight – the personal importance the researcher assigns to each factor.

Risk – additional costs such as policy changes or the possibility that the result will not be recognized.

After the CCF/CSTA statements, the career value of a NeurIPS paper for researchers inside the Chinese system drops to near zero because the paper can no longer be used as a "representative achievement". Consequently, the rational choice for most researchers becomes to target other conferences, where the expected utility remains positive.

Trust‑Capital Decay Model

NeurIPS prestige functions as a form of "trust capital" backed by authoritative lists such as the CCF recommendation list. This trust capital can be approximated by an exponential decay function: Trust(t) = Trust_0 * e^{-λt} Under normal conditions the decay rate λ is very small, allowing trust to accumulate slowly. A major coordinated event (e.g., the public statements by CCF and CSTA) can trigger a coordination effect that sharply increases λ, causing rapid decay.

In a simple coordination‑game model, let N researchers each decide independently whether to continue submitting to NeurIPS. If the proportion of "non‑submitters" exceeds a critical threshold θ, the academic reputation component for the remaining submitters declines sharply because fewer papers receive exposure and citations. This creates a phase‑transition point: if (non_submitters / N > θ) then Reputation → 0 Beyond this point, the equilibrium shifts to a self‑reinforcing non‑submission state.

Why the Apology Does Not Change the Equation

The apology only modifies attitudes, not the underlying payoff structure. As long as the CSTA rule that NeurIPS papers cannot be counted as representative work remains, the career utility for domestic researchers stays near zero, so the rational decision does not change.

From the trust‑decay perspective, the coordinated statements have already pushed the system past the critical threshold; each additional "non‑submission" further stabilizes the new equilibrium without any top‑down pressure.

Practical Implications for Researchers

Early‑career researchers are primarily concerned with concrete questions:

Can I still submit the paper before the deadline?

Which alternative venue minimizes impact on my graduation or job prospects?

Will the result be recognized for my defense or promotion?

The utility model directly answers these: switching to another top conference maximizes expected utility under the current policy environment.

The episode illustrates that academic evaluation in China depends not only on community endorsement but also on administrative definitions of what counts as a valuable output. When that definition changes, the incentive landscape shifts instantly.

Long‑term effects—such as the rise of domestic journals or a bifurcated AI research ecosystem—remain uncertain, but for researchers preparing submissions in 2026 the decisive factor is the imminent deadline, not abstract geopolitical debates.

Data sources: Chinese institutions contributed roughly 50 % of NeurIPS 2025 papers; CCF statement released 2026‑03‑25; CSTA statement released 2026‑03‑27.