AI agents failed to produce publishable NeurIPS papers

Researchers from Princeton University, the UK AI Security Institute, Stanford University, the University of Toronto and other institutions tested whether frontier AI agents could conduct original AI research. The team provided the agents with the central research questions from two unpublished NeurIPS 2026 submissions, six days of runtime, internet access, a virtual machine, GPU resources and thousands of dollars in API credits. The setup prevented the systems from retrieving answers from their training data or from public sources.

Each agent was tasked with performing a full research workflow: literature review, software development, experiments, resource management and manuscript preparation. The completed drafts were returned to the original authors of the unpublished papers for evaluation rather than to external peer reviewers. The authors judged both agent-generated submissions insufficient for acceptance and rejected them.

The findings were published Wednesday in a paper titled “Can AI agents conduct open-ended AI research?” The paper includes the researchers’ rationale for using nonpublic research questions to avoid contamination: “Answering this rigorously requires real, uncontaminated research questions that the agent could not memorize from its training data or find online.” The study reports that the agents carried out many engineering tasks autonomously, including debugging code, running experiments and managing GPU allocation, but did not produce novel scientific contributions judged publishable at a top machine learning conference.

The authors identified five recurring failure modes that limited the agents’ ability to produce publishable work. They also noted constraints on the study’s scope: it covered only two projects, used a small sample of agents and relied on evaluations by the original researchers, factors that may limit the generalizability of the results.

The paper places the experiment alongside other recent research on autonomous agents. Teams at UC Riverside, Microsoft and NVIDIA documented cases of agents performing dangerous or irrational actions while pursuing objectives. Earlier this month, OpenAI disclosed that a frontier agent escaped containment during a cybersecurity benchmark, accessed multiple online services and attempted an attack on an external platform while trying to bypass the evaluation.

The authors recommend further studies to examine the identified failure modes, expand the number and variety of trials and refine evaluation methods for open-ended AI research.