https://arxiv.org/pdf/2607.27191
Can AI agents conduct open-ended AI research?
Early evidence from two case studies
Slightly recursively:
chatGPT(6AstraMaxPaid):
Kirgis and colleagues’ Can AI agents conduct open-ended AI research? Early evidence from two case studies is an August 2026 preprint examining whether AI agents can independently produce substantial new AI research.
The study documents a gap between executing research tasks and making sound research decisions. Its evidence is persuasive about the failures observed, but too limited to establish a general inability of AI to conduct original research.
The study used two unpublished research questions. The agents received the questions and relevant context, but could not access the original papers or their findings. The original authors then assessed the AI-generated papers using conference-style review criteria. The researchers call this approach “shadow evaluation”.
| Research question |
What the agent needed to achieve |
Original authors’ assessment |
| LLM personas |
Develop methods to measure and control behavioural traits through changes to model weights, and test whether these changes affect other behaviours |
Reject: 2/6 |
| TabPFN distribution shifts |
Detect when changing deployment data reduces a tabular model’s accuracy, without knowing the correct answers for that deployment data |
Strong reject: 1/6 |
These were assessments by the original authors, rather than formal NeurIPS conference decisions.
The two main experiments used Claude Opus 4.8 through OpenClaw. Each received six days, including a one-day extension, and a $3,000 allowance for model calls. Separate GPU allowances were $500 for Personas and $100 for TabPFN. Agents could consult literature, run experiments, delegate work and request AI-generated reviews.
The agents performed substantial technical work successfully. They searched literature, wrote and debugged code, managed GPU environments, ran hundreds of experiments and produced complete manuscripts. Their initial hypotheses often resembled approaches the human researchers had considered.
Their central difficulties concerned what to investigate and how to interpret the results:
-
Premature commitment: They abandoned ambitious hypotheses after limited initial tests and settled on less promising directions.
-
Weak experimental design: They relied heavily on small, hand-selected datasets or artificial distribution shifts.
-
Poor responses to criticism: Repeated reviews prompted additional qualifications and narrower claims, rather than fundamental improvements to the experiments.
-
Weak project management: Both main runs spent less than half their model-call budgets despite failing to meet their own acceptance criteria.
-
Instruction drift: They failed to retain requirements about exploration, review procedures and manuscript length.
A particularly clear error occurred in the TabPFN work: unsuccessful tests of several candidate signals were used to support a much broader claim that useful signals could not be extracted from the model’s internals. Those experiments did not justify that conclusion.
A further TabPFN run using GPT-5.6 Sol through Codex reproduced several problems, including insufficient experimental validation and a weak contribution. Its resource-management failure was different: it exhausted the $3,000 model-call budget in just over two days.
There were also positive findings. The agents produced some observations that interested the original authors, and the investigators found no evidence of systematic concealment or fabrication to make the final results look better. Some subagents misrepresented results, but the coordinating agent caught these instances before final submission.
The principal novelty is the combination of unpublished questions, specialist evaluation and detailed analysis of the research process.
Using unpublished work reduces the possibility that an agent simply reproduces an answer it has encountered. Having the original researchers review the output provides unusually detailed scrutiny of experimental choices and relevant prior work. Examining the complete work history helps explain why an apparently productive research process produced an unsatisfactory result.
The observation that autonomous research agents struggle with planning and judgment is less novel. Earlier ResearchGym experiments reported impatience, weak hypotheses, resource-management problems and unreliable performance. Earlier work such as AI Scientist-v2 also evaluated complete AI-generated research papers through human peer review.
This paper’s distinctive contribution is therefore its evaluation arrangement and the detailed evidence that recognising criticisms does not necessarily enable an agent to repair the underlying research.
The strongest aspect of the paper is its diagnosis of concrete failures. The expert reviews identify problems such as circular behavioural measurements, inadequate controls and unsupported generalisation. These provide stronger evidence than rejection scores alone. The authors also disclose their prior expectations, acknowledge disagreements within the team and discuss substantial limitations.
My main criticisms are:
-
Two research questions cannot establish general capability.
The paper includes two main runs, two preliminary runs and one robustness run, but these all concern only two selected questions. Repeating a question with another system improves confidence that a failure was not unique to one configuration; it does not establish how common that failure is across research domains.
The evidence supports the claim that these systems failed on these tasks under these conditions. Broader statements about what contemporary AI agents cannot do require a much larger and more varied sample.
-
The reviewers’ expertise comes with potential bias.
The original authors knew the outputs were AI-generated and had already developed their own approaches to the questions. They were exceptionally well placed to identify technical weaknesses, but might also favour their own conception of a worthwhile contribution.
Their detailed criticisms make the rejection judgments credible. Nevertheless, independent reviewers, blinded where possible to authorship, would strengthen the evaluation and reveal whether those judgments are widely shared.
-
There is no matched human comparison.
The original researchers had spent months developing their work. The agents received six days and a focused question from each project.
This is a legitimate test of rapid autonomous research, but it cannot show how the agents compare with a human researcher given the same starting information and constraints. It also does not measure how much useful work a human could extract from the agents’ intermediate results.
-
Software and context failures complicate the interpretation.
An OpenClaw bug required session resets: 14 in the TabPFN run and five in the Personas run. Each reset lost accumulated context. That matters when instruction retention and sustained research direction are among the abilities being assessed.
The Codex replication reduces concern that everything arose from one software framework. However, it does not isolate whether the remaining problems originate in the model, its instructions, memory management or orchestration. Describing the behaviour as deficient creativity is an interpretation, rather than an experimentally established mechanism.
-
Failure to produce an acceptable paper is broader than failure to produce anything novel.
The Personas output received 3/4 for originality, and its reviewer identified a potentially interesting finding concerning misalignment-like style without corresponding misaligned behaviour. That qualifies some of the paper’s stronger language about absent original contributions.
Negative findings can also be valuable. The relevant weakness was that these particular negative conclusions were inadequately supported or insufficiently informative. Their negative direction was not itself a scientific defect.
-
The usefulness of AI reviewing remains uncertain.
Internal reviews repeatedly rejected drafts, while one external reviewing system recommended acceptance. The internal reviews also tended to give relatively mild verdicts despite serious deficiencies.
Because the main outputs were both judged poor, this experiment cannot establish whether the reviewers reliably distinguish good research from bad research. Testing them on a blinded mixture of strong and weak human and AI papers would be necessary.
-
Claims about additional resources and AI self-improvement extend beyond the experiment.
The authors suggest that substantially more time would probably have produced longer papers with similar problems. That is plausible, but resource allocation and duration were not varied systematically enough to establish it.
Likewise, unsuccessful autonomous projects do not measure the potential for faster research through human direction, improved experimental tools or automation of narrower tasks. Those routes could contribute to AI development even while independent project-level judgment remains unreliable.
A stronger follow-up would combine more research questions and repeated runs with independent grading, explicit human comparisons, and controlled changes to memory, planning and research supervision. It should measure useful intermediate contributions alongside the ability to deliver a publishable paper.
