Step 10: tool results

read_page

{
  "url": "https://arxiv.org/html/2607.27191v2",
  "offset": 66000,
  "max_chars": 12000
}

{"ok":true,"url":"https://arxiv.org/html/2607.27191v2","text":"s\nstopped early, the loop resumes after another GPT-5.6 Sol model reads the PDF\nand completes a NeurIPS review, which is then passed back to the agent.\nIn this setting, we observed many of the same failure modes. The agent failed\nto run appropriately powered experiments, did not make a novel contribution,\nand returned a draft with misformatted figures and no appendices.\nIt also failed to manage its budgets, although not in the same way as our\nOpenClaw experiments. Given the same budget for token usage, GPT-5.6 exhausted the\n$3,000 budget in just over two days, leaving nearly 100 hours left of the\nallotted time. This partially explains the underpowered experiments; the agent\nspent the majority of the time iterating on different hypotheses and only\nbegan scaling a candidate solution after depleting the majority of its token\nusage budget. It then had only a few hundred dollars for paper writing, which\nit consumed in only a couple of iterations.\nThe experiment also reproduced some of the positive findings. The AI\nself-review process continued to appropriately return rejects on the drafts\nthe agent produced. The agent followed good scientific ethics, registering its\nhypotheses and reporting a negative finding rather than inventing a positive\nresult. It also made one significant improvement over our first experiments.\nUnlike the OpenClaw experiments, which almost exclusively used synthetically-shifted data sources, the\nagent found and worked with a real-world distribution-shifted dataset.\n7 Limitations\n7.1 Elicitation threats\nWe ran our main experiments using OpenClaw since we wanted to be able to switch\nbetween model providers. In an early pilot run, we tested GPT-5.3 Codex, and\nswitched to Opus 4.8 after recognizing that GPT-5.3 Codex could not effectively\nuse this scaffold. While we ran a follow-up experiment with GPT-5.6 Sol using\nCodex to assess the robustness of our results—that is, to check that we were not\nsignificantly under-eliciting performance relative to a naive implementation\nin the default harness—we did not devote the same time to harness engineering\nin that scaffold as we did for our main experiments, which had multiple\niterations of scaffold refinements.\nWe think using vendor-provided scaffolds would also improve the reliability of\nthe scaffold. Partway through our runs, we found that OpenClaw’s agent loop\nconflicts with the cryptographic signatures that Anthropic attaches to its\nthinking blocks, and the conflict crashes the session. This was an open\nOpenClaw issue during our experiments.\nTo fix the bug, we modified the scaffold to reset the session and point the\nagent back to its project files with a short summary of the error. This reset\nwas triggered 14 times in the TabPFN run and five times in the Personas run.\nEach reset cost the agent accumulated context. We do not think this\nmeaningfully impacted our results: the two runs differed sharply in how often\nthey encountered this bug, but they did not differ in the quality of the final\npaper or the failure modes we encountered. Still, we expect vendor-provided\nscaffolds to be better suited to running their models and we do not expect such\nreliability issues to arise in these scaffolds.\nWhile current evidence suggests capable open-source scaffolds are within the\nmargin of error on long-horizon tasks, two-thirds of our respondents said a\nfailed run might be explained by scaffold limitations.\nFinally, the agents had six days to complete the experiment. On one hand, this\nis longer than most existing evaluations of automated AI research. On the other,\nthe original authors spent much longer on their paper, and their training runs\nconsumed far more GPU hours.\nThere are two reasons we do not think this meaningfully impacted our results.\nFirst, neither agent fully used its computational resources, and the expert\nreviewers’ main objections were about the quality of experiment choice,\njudgment on data selection, and poor reasoning about the negative AI reviews,\nnot the quantity of experiments.\nSecond, we decided the compute budgets based on authors’ estimates of how much\ncompute would allow agents to answer a specific research question from their\npaper. In particular, we chose just one research question from their original\npaper, and the agent was still unable to make progress towards answering it.\nAll of these reasons lead us to believe that giving the agents more time would\nhave produced a longer paper with the same central limitations; it would not meaningfully change the main results.\nFinally, we could not test Anthropic’s strongest model. Anthropic deliberately\nlimited Fable 5’s abilities on frontier AI R&D. We are working to get access\nto Fable/Mythos 5 for future experiments.\nAt the same time, there is also a strong contrast between our findings for our\nautomated AI research experiments and our previous experiment on iOS app development. In our\nprevious experiment, a similar agent setup—OpenClaw with Opus 4.6\nthinking—autonomously built and shipped an iOS app  ( Kapoor et al., 2026 ) . In other\nwork, we have found that frontier models using open-source agent scaffolds can\nnow reproduce published research far better than they could two years\nago  ( Nadgir et al., 2026 ) . This suggests that unlike engineering tasks and\nverifiable research tasks, agents struggle with some kinds of open-ended AI\nresearch tasks.\n7.2 Implications for accelerating AI research\nIn this paper, we claim to find preliminary evidence that AI agents are not yet capable of conducting open-ended autonomous AI research. We highlight numerous identification challenges with shadow evaluations and the particular experiments that we run. But our broader motivation for this work is to evaluate claims of AI agents accelerating and even automating AI R&D itself. Thus, a separate question concerns the implications of a positive or negative finding on the question of AI agents automating open-ended AI research.\nThere are multiple reasons why our results might not actually clarify the debate on the broader question of accelerating AI R&D. It is possible that the path to AI R&D does not require full automation of open-ended tasks like the ones we study, or that the open-ended research skills that we measure are not ones on the critical path.\nNevertheless, there are good reasons to view the distinction between AI capabilities on verifiable tasks and open-ended research problems as germane to the question of accelerating AI R&D. Anthropic’s post on self-improvement explicitly cites Claude’s rising success rate on LLM-judged \"open-ended\" Claude Code sessions as evidence of self-improvement  ( Favaro and Clark, 2026 ) . A recent report from the Elasticity Institute on the economics of recursive self-improvement explicitly distinguishes between \"broad\" and \"narrow\" AI capabilities, discusses implications of a speed-up in only \"narrow\" capabilities, and cites the need for more data on the full breadth of AI capabilities and weaknesses relative to human AI researchers  ( Cunningham et al., 2026 ) . We hope our experiments provide early evidence to answer this question.\n8 Potential biases\nOpen-world evaluations allow substantial researcher discretion:\nresearchers have leeway in choosing the questions being evaluated, designing\nthe study, and executing it. This means that our prior beliefs, expectations,\nand biases—especially those of the core team that designed and conducted the\nexperiments—could impact the results of our evaluation. In particular, some\nof the core team members are known for our position that imminent recursive\nself-improvement leading to runaway superintelligence is unlikely; this could\naffect how we design the evaluation and how we interpret the results.\nFor example, we describe the agents’ failures as the lack of creativity and\njudgment: the agents settled into their initial approach too quickly when the\ntasks required creative problem solving. As we discuss in the text, this\ninterpretation is not self-evident, and some of our coauthors instead interpret\nthe results as failures of reasoning and logic, or as epistemic lock-in.\nSimilarly, the studies we chose for this evaluation reflect our understanding\nof what constitutes results of an empirical NeurIPS paper; but other\nresearchers might disagree on the level of open-endedness that is necessary\nfor making progress towards recursive self-improvement.\nGiven the nature of the evaluations, we do not think there is an “unbiased”\nway to conduct them. While benchmarks with clear success criteria are in some\nsense more objective, that also leads to a narrower task specification and\ncurtails the types of research questions that can be studied; open-world\nevaluations trade off objectivity for a much richer set of evaluation tasks.\nThese considerations suggest several principles for designing open-world evaluations. We think such evaluations are most persuasive when epistemically diverse sets of people conduct them. This includes both within-team and across-team diversity in ideas. It is also important to disclose the team’s biases and positions and explicitly surface disagreements. As an example, many coauthors have different priors compared to the core team; this helped us surface disagreements in our interpretation of the results when they did occur.\nWe also took many steps to minimize the effects of the core team’s priors.\nThis included running multiple dry runs focused on improving the scaffold,\nallowing the agent generous budgets, and analyzing the logs to understand and\naddress key failure modes. For example, we added multiple lines of prompting\nto address some recurring failure modes, such as the lack of exploration.\nIn addition to these “known” biases, we are also subject to “unknown”\nbiases that are hard to identify a priori. For example, many of our results\nrely on analyzing the logs of the agent. Log analysis involves discretion, and\nwe may have identified failures that fit our expectations more readily than\nones that did not. We release the full expert reviews, survey responses, agent\nrepositories, and run logs so readers can check our interpretation against the\nraw materials.\nFinally, we plan to continue evaluating AI R&D automation using new research, and we welcome feedback and adversarial collaborations for follow-up studies.\nAuthor contributions\nCore team: Sayash Kapoor and Arvind Narayanan conceptualized the\nproject. Peter Kirgis and Andrew Schwartz implemented the agents, led the log\nanalysis, and designed the website. Sayash Kapoor and Peter Kirgis drafted the\npaper. Stephan Rabanser, as an author of one of the original papers, reviewed\nthe papers produced by the agents. All members of the core team (Peter Kirgis,\nSayash Kapoor, Andrew Schwartz, Stephan Rabanser, and Arvind Narayanan)\ncontributed to editing and responding to feedback.\nLog analysis: Matilda Orona, Tilman Bayer, Derrick Chan-Sew, Yue Ling,\nAbhishek Shetty, Toby Pilditch, and Nitya Nadgir contributed to the log\nanalysis.\nCollaborators: Cozmin Ududec and Magda Dubois contributed to the\nconceptualization of the project. David Africa, Konstantinos Voudouris, and\nViet Nguyen, as authors of the original papers, reviewed the papers produced by\nthe agents. Cozmin Ududec, Magda Dubois, David Africa, Konstantinos Voudouris,\nToby Pilditch, Harry Coppock, Helen Toner, Gillian Hadfield, Seth Lazar, Steve\nNewman, and Shoshannah Tekofsky responded to the pre-experiment survey and,\ntogether with Rishi Bommasani, offered feedback and inputs into the essay text\nand our analysis and interpretation of the results.\nAcknowledgments. We are grateful to Andy Hall for participating in the survey. We thank Mariia Koroliuk and Anton Gonzalvez Hawthorne for helpful comments on one of the LLM generated papers.\nAI disclosure. We used AI for converting the original draft of the paper from a Google doc to L a T e X , converting links to references, copyediting, as coding assistants for developing the agent and writing documentation, conducting AI-ass","chars":12000,"total_chars":124208,"offset":66000,"truncated":true,"next_offset":78000}