ML research is not serious research, NeurIPS board statement suggests (statmodeling.stat.columbia.edu)

The NeurIPS board recently addressed the rising concern over "hallucinated references" in machine learning (ML) papers presented at the 2025 conference, revealing that at least 53 papers included citations that could not be verified. This issue raises critical questions about the conference's policy on handling papers with such inaccuracies—whether they should be retracted, corrected post-submission, or subjected to stricter review processes. The board acknowledged the ongoing evolution of large language models (LLMs) in the research community and the complexities they introduce, but maintained that the presence of hallucinated references does not inherently invalidate a paper's content. The significance of this statement lies in its implications for the integrity of research in the AI/ML field. While some argue that these inaccuracies might stem from honest mistakes, the board's position suggests a concerning tolerance for fictitious evidence, risking the credibility of scholarly work. It highlights an emerging dilemma for researchers: balancing the efficiency gained from LLMs with the necessity of rigorous citation practices. The NeurIPS board's ongoing deliberation on this issue reflects a broader uncertainty in the community regarding the reliability of AI-generated content, and underscores the pressing need for clearer standards in the era of automated writing assistance.