Claims that AI agents are close to doing autonomous AI research now face a pointed challenge. A new paper from Princeton and the UK AI Security Institute tested whether frontier models could move from a research question to a publishable result, using unpublished NeurIPS papers as the benchmark.
The outcome was not that the agents were useless. They could search literature, debug code, run experiments, and assemble LaTeX papers. The deeper problem was that they did not show the research judgment needed to produce work that expert reviewers considered meaningful.
A Shadow Evaluation for AI Research
The researchers used an approach they call "Shadow Evaluation." Instead of asking an agent to solve a narrow task with an easy-to-check answer, they gave it the central research question from an unpublished paper. The original authors, who had worked on the same question for months, then evaluated the AI-generated result as if they were conference reviewers.
This setup matters because the papers were not yet on the web. The agent could not simply rely on material that might already exist in training data. It had to plan, experiment, interpret results, and write up the work under conditions closer to open-ended research.
The team worked with the authors of two NeurIPS 2026 submissions. One studied how personality traits of language models can be steered through their weights. The other developed a method called TabPFN for detecting when a tabular prediction model encounters deployment data that differs sharply from its training data and causes accuracy to collapse.
The main runs used Claude Opus 4.8 with Extra-High Reasoning. Each agent had six days, $3,000 in API credits, a GPU budget, a virtual machine, and access to the open web. The agents ran inside a scaffold that could orchestrate model calls, launch subagents, monitor resources, and use external AI review tools.
The Papers Failed Expert Review
The finished papers did not pass the judgment of the original authors. Both were rejected, and one received a "Strong Reject." Reviewers pointed to weak motivation, poor experiment choices, difficult prose, and a lack of new contributions.
The criticism went beyond surface polish. One reviewer described the reasoning as a "'proof by example' fallacy" and "highly non-scientific." Another said the experiment choices were "bizarre" and that the results appeared to come from "post hoc choices."
The logs showed recurring weaknesses. The agents could generate plausible hypotheses, but they did not reliably judge what would meet the standard for publishable research. They often leaned on small, hand-curated, or synthetic datasets, then discarded ideas without building a stronger research direction.
Creative recovery was also weak. When early hypotheses failed, the agents tended to narrow their claims instead of finding more promising paths. Their internal AI reviewers never gave a single Accept across fifteen revision rounds, yet the agents did not fix the central objections.
Engineering Was Stronger Than Judgment
The agents did complete a large amount of technical work. They ran literature searches, debugged GPU code, completed hundreds of experiments and robustness tests, and produced full papers in LaTeX. Only three human interventions were needed: a scaffold bug fix, a deadline extension, and a request to rewrite for readability.
That distinction is central to the study. The agents were competent at many research engineering tasks, but research is not only execution. It also requires choosing the right question, knowing when evidence is thin, recognizing when a line of work is not adding enough, and deciding how to respond to criticism.
The agents also struggled with backtracking and time management. Both abandoned their most ambitious research goals within the first ten hours. In the Personas run, the exploration phase ended after just five hours, even though the plan had set aside 36 to 48 hours.
Resource use created another warning sign. Both runs finished with less than half the API budget spent. One agent marked the project complete seven hours before the deadline, shortly after its own reviewer had returned another Reject.
Long-running instruction follow-through was also fragile. The agents drifted away from explicit requirements over time. Both papers exceeded length limits and would have been desk-rejected at NeurIPS. One paper had zero visualizations in the main text, while the human-written original had 15 figures.
Why More Compute May Not Solve It
The researchers checked whether the software environment itself was causing the failures. They repeated one experiment with GPT-5.6 Sol and OpenAI's Codex scaffold. Nearly all the same problems appeared, while GPT-5.6 used the $3,000 budget in just over two days and ended up with undersized experiments.
The study did find that more reasoning effort improved quality compared with preliminary tests without reasoning. But the researchers doubt that simply adding more time or compute would change the core result. The expert complaints focused on the quality of the experimental choices, not on the number of experiments.
The authors say frontier models can handle the engineering side of AI research but "cannot solve weeks-long, open-ended AI research questions." They also note limits to the study: it covered only two papers, and the reviewers were not blinded. The reviewers knew their own research question and knew the work was AI-generated.
Even with those limits, the results challenge a common shortcut in evaluating autonomous research systems. Previous claims have leaned heavily on accepted submissions, but acceptance alone may not say much about research quality. Sakana AI's The AI Scientist-v2 submitted three papers to an ICLR workshop in 2025; one was accepted with an average score of 6.33, barely above the threshold, then withdrawn after researchers found citation errors.
Workshop acceptance rates also differ from main conference standards. The source notes 60 to 70 percent for workshops, compared with 20 to 30 percent at main conferences. Shadow Evaluation tries to avoid that ambiguity by having the original experts evaluate the agent against the work they actually did.
A Clearer Line Around Autonomous Agents
The study does not show that AI agents are irrelevant to research. It shows a more specific boundary. They can automate parts of the workflow that look like engineering: running jobs, fixing code, collecting results, and preparing a manuscript.
But the same agents failed at the parts that decide whether the work is strong research. They did not reliably choose convincing experiments, respond to deep criticism, preserve important instructions, or use their available time and budget well.
That distinction cuts directly against claims from leading AI labs. Anthropic published "When AI Builds Itself" in June, discussing internal data on research acceleration and raising the idea of a globally coordinated development pause. OpenAI said GPT-5.6 Sol helped with post-training a smaller model and saved researchers several weeks, while the study's authors note that this contribution is not mentioned in the 81-page system card.
The researchers have made expert reviews, logs, and agent repositories available online so others can inspect the evidence. For now, the clearest conclusion is narrow but important: AI agents can assist research engineering, but this experiment did not show that they can independently produce strong, open-ended AI research.