Overview
A team of researchers from Princeton, Stanford, MIT, and other institutions—including Helen Toner and Arvind Narayanan—published a 2026 study, *Can AI agents conduct open-ended AI research? Early evidence from two case studies* (arXiv: 2607.27191), introducing an evaluation method called Shadow Evaluations.
The premise: give an AI agent the same conditions as a graduate student—six days, a few thousand dollars of compute, a computer—to tackle the core research question of a real top-conference submission, then see if it can reproduce genuine research progress.
The Problem with Existing Evaluation Methods
Current approaches sit at two extremes:
- Narrow task benchmarks (e.g., "raise MNIST accuracy to 99%") measure engineering skill, not research ability—real research is navigation through fog, not point-A-to-B routing.
- Peer review of AI-written papers suffers from inconsistent, noisy reviewing.
- Kirgis, P., et al. (2026). *Can AI agents conduct open-ended AI research? Early evidence from two case studies*. arXiv:2607.27191.
How Shadow Evaluations Work
The researchers extracted the core research questions from two high-quality, unpublished NeurIPS 2026 submissions. AI agents were given the questions, six days, and a compute budget with no human intervention. Crucially, the original authors of each paper judged the AI's output—they best understand what the research question is, where the difficulties lie, and what a valid answer looks like.
Results: Engineering Giant, Research Dwarf
The agents completed all engineering work—building experimental frameworks, implementing baselines, running large numbers of experiments, producing plausible charts and analysis. But on both papers' core research questions, they made no substantive progress. The original authors said both AI "papers" would be immediately rejected.
Five Failure Modes
1. Misjudging the publication bar — treating trivial results (e.g., a 2% improvement) as breakthroughs without asking whether they're statistically significant, generalizable, or explainable; lacking the trained "taste" for what counts as a meaningful finding. 2. No creative response to design flaws — when results diverge from expectations, the agent tweaks parameters within the existing frame rather than questioning whether the approach itself is wrong. 3. Persisting in dead ends — like a stubborn gambler, it keeps investing in clearly failing directions instead of making a strategic retreat; it can't judge when to persist and when to abandon. 4. Weak resource awareness — despite budget limits, it launches many expensive experiments in parallel without prioritizing which are most valuable; it lacks a sense of cost. 5. Instruction drift — over time it gradually wanders away from the core research question, doing interesting-but-irrelevant work without any self-check mechanism.
What This Means
The results are a warning that optimistic forecasts of AI automating research within a few years need recalibration. Agents can autonomously complete engineering—a genuine achievement—but research also demands judgment, creativity, strategic thinking, resource management, and self-reflection, which are exactly where current AI is weakest.
Limitations and Outlook
The study covers only two case studies with a specific model and scaffold; robustness checks with a different model and scaffold produced consistent results, but this remains early evidence. Shadow evaluations nonetheless provide a template for future assessment: have the people who understand the problem best judge the quality of the answer. The authors conclude that writing code and doing research are two different things—just as writing words and writing poetry are.
> Note: NeurIPS (Neural Information Processing Systems) is one of the most selective machine learning conferences; publishing there is generally considered a marker of research excellence.