The pitch from Anthropic and OpenAI has been simple: their models can accelerate AI research, maybe even automate it. A new experiment from Princeton and the UK AI Security Institute suggests that’s premature. The team tested whether frontier agents could replicate unpublished NeurIPS submissions using only the core research question. Both attempts were rejected by the original paper authors acting as reviewers. One earned a “Strong Reject.”
The setup was generous by any standard. Claude Opus 4.8 with Extra-High Reasoning got six days, $3,000 in API credits, GPU access, a virtual machine, and the open web. The agent ran in OpenClaw, an open-source scaffold built by Peter Steinberger before he joined OpenAI. It could spawn subagents, monitor its own resources, and consult external AI reviewers. The researchers called this “Shadow Evaluation” — because the papers weren’t yet public, the agent couldn’t simply retrieve answers from training data.
The engineering worked. Agents ran literature searches, debugged GPU code, completed hundreds of experiments, and compiled LaTeX papers with essentially no human help. But the research judgment collapsed. Reviewers flagged unreadable prose, poorly motivated experiments, and a “‘proof by example’ fallacy” they called “highly non-scientific.” The agents abandoned their most ambitious goals within ten hours, spent less than half their API budget, and one declared the project complete seven hours early — right after its own reviewer issued another Reject. Fifteen internal review rounds produced zero Accepts, yet the agents never addressed the core criticism.
This isn’t just a benchmark miss. Meta AI recently described a similar phenomenon called “behavioral state decay,” where agents forget explicit requirements over long contexts. Both papers exceeded NeurIPS length limits and would have been desk-rejected. One contained zero figures; the human original had 15. The pattern across frontier models suggests the bottleneck isn’t compute or capability, but something closer to taste — knowing what a publishable contribution looks like, and having the conviction to pursue it when the first hypothesis fails. That’s a much harder problem to solve than scaling GPUs.
