A research team from multiple institutions, led by Peter Kirgis and Sayash Kapoor at Princeton University, tested whether current AI agents can carry out open-ended AI research and concluded they are not yet able to produce original work at the level expected by top machine-learning conferences.
What the researchers did
The authors introduced a new evaluation method called “shadow evaluation”: agents had to answer research questions drawn from high-quality, unpublished papers. The experiment used two submissions to the NeurIPS 2026 conference; because the papers were not public, the agents could not rely on memorized or online sources.
Anthropic’s Claude Opus 4.8, run on the open-source OpenClaw framework, was assigned the tasks. Each agent had six days, $3,000 in Anthropic API credits, a GPU budget, virtual machines, and access to the open web to produce a paper worthy of a top-tier conference. The original paper authors graded the agent-generated manuscripts as if they were peer reviews for a conference submission.
The research problems
One task asked whether a large language model’s ‘‘personas’’—the behavioral tendencies produced by model weights—can be controlled by editing the weights themselves. The other task concerned designing a detector to flag when a model that makes predictions from spreadsheet data has become unreliable.
Findings: engineering succeeds, creative research fails
The original authors rejected both AI-generated papers. Reviewers concluded that agents could perform the engineering components of research: they reviewed literature, ran hundreds of experiments, and collected results. However, the systems failed to produce the kind of creative, judgment-based work necessary for an original, publishable contribution.
According to Kapoor, the agents were “unambiguously bad” at carrying out the research itself: they ran odd or weak experiments (for example testing hypotheses on tiny synthetic datasets), struggled to write intelligibly about their work, and did not make novel contributions. The agents tended to commit too early to unpromising approaches, did not sufficiently explore alternatives, and lacked the ability to fundamentally rethink failed directions—only small pivots were possible.
Resource use and feedback
Agents poorly managed resources such as tokens, compute time, and task scheduling; they failed to follow instructions about time allocation for research phases or constraints on paper length. They also did not effectively incorporate feedback from subagents or external AI reviewing tools. Instead of revising methods, agents narrowed claims and added caveats.
A notable positive finding was the absence of ‘‘reward hacking’’—the main orchestrator did not hide or falsify experiments or data. Subagents occasionally hallucinated or misreported results, but the orchestrator typically detected those issues.
Why this may be happening
Kapoor attributes the gap to how models are trained. Current training regimes, such as reinforcement learning, are easier to apply to tasks with automatically checkable success conditions. It is more difficult to create training environments for genuinely open-ended tasks that require creativity and nuanced judgment.
The team is repeating the experiment with Mythos, Anthropic’s most advanced model, which launched in April. (Mythos is now subject to U.S. government safety restrictions and available only to approved organizations.) Anthropic did not respond to a request for comment.
Limitations of the study
The study has limitations: it examined only two papers, and the original authors knew they were grading AI-generated work, which might have influenced their evaluations. Researchers also had considerable discretion in designing and running the experiment, allowing prior beliefs or biases to affect outcomes. Evaluating open-ended research trades some objectivity for a richer, more realistic test than standard benchmarks provide.
Implications for recursive self-improvement timelines
The results temper expectations that recursive self-improvement—AI systems rapidly improving themselves with little human oversight—is imminent. Industry players have signaled progress toward automation: Anthropic published a June blog post titled “When AI Builds Itself,” and in July OpenAI noted that GPT-5.6 Sol helped post-train a smaller model, saving researchers weeks.
Anthropic cofounder Jack Clark has written that today’s systems are excellent engineers but often lack ‘‘valuable, intuitive creativity,’’ a short-term bearish signal for recursive self-improvement. Najoung Kim, a Boston University professor who researches AI automation of research, said targeted investment and effort could produce progress, but it is also possible that advancement will split: rapid gains on narrow, scorable tasks, and slower progress on open-ended research.
The core open question is how essential creative, open-ended research is to achieving recursive self-improvement. Kapoor notes that major past advances in AI—such as the invention of the transformer architecture—required creative leaps. Others hypothesize that incremental gains (speeding up training, improving benchmarks) might suffice. As Kapoor put it, that is currently the “trillion-dollar question.”
