Anthropic and OpenAI have been touting their models' ability to speed up AI research. A new experiment using unpublished NeurIPS papers tells a different story.

Can AI agents conduct AI research on their own? A new paper from Princeton and the UK AI Security Institute puts that claim to the test. Today's frontier models can handle research engineering, but they fail at the parts of the research process that actually matter.

Solid evidence for claims about automated AI research has been mostly absent, the authors argue. Existing evaluations either test agents on narrow, verifiable tasks or submit AI-generated papers to peer review, a process the researchers call "overstretched, stochastic, and suffers from poor review quality."

A new way to test AI research ability

The researchers call their approach "Shadow Evaluation." An agent receives the core research question from an unpublished paper. The original authors, who spent months on the same question, then evaluate the result as conference reviewers would. Since the results aren't on the web yet, the agent can't fall back on training data.

The team partnered with the authors of two NeurIPS 2026 submissions. The first paper examines how personality traits of language models can be steered through their weights. The second develops a method called TabPFN that detects when a tabular prediction model hits deployment data that differs sharply from its training data and tanks its accuracy.

The main experiments used Claude Opus 4.8 with Extra-High Reasoning. Each agent got six days, $3,000 in API credits, a GPU budget, and full access to a virtual machine and the open web. The agent ran inside a scaffold, a software environment that orchestrates model calls and provides tools, letting the model delegate to subagents and monitor its own resource usage. It could also consult external AI review tools.

The researchers used OpenClaw, an open-source, vendor-neutral agent framework built by Austrian developer Peter Steinberger, who joined OpenAI earlier this year. The core agent launches subagents and long-running GPU jobs, with an automatic heartbeat from the scaffold waking it up when jobs finish so it can collect results.

The original authors reviewed the finished papers as conference reviewers and rejected both. One got a "Strong Reject." They criticized poorly motivated data and experiments, unreadable prose, and no new contributions. One reviewer called the reasoning a "'proof by example' fallacy" that was "highly non-scientific." Another called the experiment choices "bizarre" and the results clearly the product of "post hoc choices."

Failure modes that keep showing up

Analysis of the agent logs reveals systematic weaknesses. The agents lacked judgment about what meets the bar for publishable research, generating plausible hypotheses but discarding them based on small, hand-curated, or synthetic datasets.

They also failed at creative problem-solving. When initial hypotheses were falsified, the agents narrowed existing claims rather than pursuing new directions. Their internal AI reviews never returned a single Accept across fifteen revision rounds, yet the agents never addressed the core criticism.

The agents couldn't backtrack effectively either. Both gave up their most ambitious research goals within the first ten hours. In the Personas run, the agent finished its exploration after just five hours despite budgeting 36 to 48 hours for that phase.

Resource awareness was poor too. Both runs ended with less than half the API budget spent. One agent declared the project complete seven hours before the deadline, shortly after its own reviewer had returned yet another Reject.

The agents also suffered from instruction drift, gradually forgetting explicit instructions over a long context. Both papers exceeded length limits and would have been desk-rejected at NeurIPS. One paper contained zero visualizations in the main text, while the human-written original had 15 figures. Meta AI recently described a closely related phenomenon called "behavioral state decay", where an agent recognizes a requirement early on but violates it later while fixing an unrelated bug.

The engineering works, the research judgment doesn't

The agents managed all engineering work without human help. They ran literature searches, debugged GPU code, completed hundreds of experiments and robustness tests, and compiled full papers in LaTeX. Only three human interventions were needed: a scaffold bug fix, a deadline extension, and a request to rewrite for readability.

The researchers found no significant reward hacking. The agents didn't manipulate results or skew data to chase better scores. In fact, they did the opposite, starting with ambitious claims and correcting them toward negative results.

To check whether the results were artifacts of the software setup, the researchers repeated one experiment with GPT-5.6 Sol and OpenAI's Codex scaffold. Nearly all the same failure modes showed up. GPT-5.6 burned through the $3,000 budget in just over two days, leading to undersized experiments.

Results clash with lab claims

The findings contrast with claims from leading AI labs. In June, Anthropic published a post titled "When AI Builds Itself", sharing internal data on its own research acceleration and floating the idea of a globally coordinated development pause.

OpenAI claimed that GPT-5.6 Sol helped with post-training a smaller model and saved researchers several weeks. The study's authors note this contribution isn't even mentioned in the 81-page system card.

More reasoning effort did improve quality, as preliminary tests without reasoning showed. But the researchers suspect more time or compute wouldn't change the results much. The reviewer's objections targeted the quality of experiment choices, not the quantity.

The authors say frontier models can handle the engineering side of AI research but "cannot solve weeks-long, open-ended AI research questions." The study only covers two papers, and the reviewers weren't blinded, knowing both their own research question and that they were evaluating AI-generated work. But the results were so clearly weak that these limitations are unlikely to change the overall picture. The team has made expert reviews, logs, and agent repositories available online so other researchers can judge for themselves.

Why peer review is a poor yardstick

Previous claims of autonomous research success relied almost entirely on accepted submissions, but acceptance alone says little 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, and withdrawn after review when the researchers found citation errors. Workshop acceptance rates run 60 to 70 percent, well above the 20 to 30 percent at main conferences. Shadow Evaluation sidesteps this problem by handing evaluation back to the original authors.

Sakana AI has since founded a dedicated research lab for recursive self-improvement in June 2026. A later version of the AI Scientist wrote a paper that passed peer review, according to the company. The related research was published in Nature in March 2026.

The search for AI that creates new knowledge

Criticism is growing alongside these empirical results. Tom Zahavy of Google Deepmind argues in the position paper "LLMs can't jump" that language models lack the cognitive mechanism to create something genuinely new. The AI Scientist merely recombines existing concepts, he writes, while DeepMind's AlphaEvolve excels at optimization but needs a clear error signal to work.

Recent math breakthroughs might seem to contradict that view, but they likely don't. In May, an OpenAI reasoning model disproved a conjecture about unit-distance geometry that Paul Erdős posed in 1946. Fields Medalist Tim Gowers called the result a "milestone in AI mathematics." Shortly after, Claude Mythos solved the same problem with what Anthropic engineer Sholto Douglas called a "cute, simple proof." OpenAI then confirmed Astra, a new model family that reportedly solved ten open problems in math and theoretical computer science where the field had been stuck for at least a decade.

But mathematical proofs are deduction, deriving necessary conclusions from fixed rules or systematically searching large solution spaces. That's impressive, but it's what frontier models are built to do. Research also requires abduction, the creative leap that invents a cause to explain a surprising phenomenon. Formulating new research questions, designing the right experiments, and drawing solid conclusions from the results are things today's models still can't do.

Read on for the full picture.
Subscribe for hype-free coverage.

• Full access to every article on THE DECODER
• No ads
• Join the comments and community discussions
• A weekly AI news recap via mail
• 6x/year: "AI Radar" — deep dives on the AI topics that matter most
• Daily AI news, always up to date
• Our full ten-year archive
• Covered by a team with 10+ years in AI