Study: AI agents still fail at real research

What this is about

An arXiv study submitted on July 29, 2026 asks a simple but hard question: can today's AI agents conduct open-ended AI research when the task is not just tests, code, and benchmarks, but real scientific judgment?

The author group, including Peter Kirgis, Sayash Kapoor, Rishi Bommasani, and Arvind Narayanan, calls the method “shadow evaluations.” An agent receives the central research question of a high-quality unpublished paper. The original authors then judge whether the agent produced a publishable answer. The result is sobering: the agents completed the engineering work without human help, but experts clearly rejected both outputs.

What the shadow evaluation actually does

Typical agent benchmarks often test tasks with clear goals: pass the test, find the file, solve the ticket. Open-ended research is different. A system must judge whether an idea is promising, which experiments matter, when a line of work is dead, and what the bar for a NeurIPS paper actually is.

In the study, frontier agents worked on two unpublished NeurIPS 2026 submissions. They were given six days and thousands of dollars of compute. The agents built repositories, ran experiments, and executed technical work. What they lacked was the ability to substantially advance the actual research question. The authors identify five recurring weaknesses: poor judgment about publishable quality, uncreative responses to design problems, weak backtracking from dead ends, poor resource awareness, and instruction drift.

Why it matters

Many forecasts of extremely rapid AI progress depend on one assumption: if agents automate AI research, progress starts accelerating itself. This study directly tests that assumption. It does not say automation is impossible. It shows that today's agents do not simply replace what strong research groups do.

For companies and developers, the finding is useful because it draws a sharper boundary. Agents are strong when goals are clear, environments are measurable, and feedback arrives quickly. Research, product strategy, and architecture decisions contain many open trade-offs. Blind automation in those settings can produce lots of activity but little insight.

In plain language

Imagine giving someone a kitchen, ingredients, and six days to invent a new bread. They can use the oven, knead dough, and keep careful tables. In the end, what matters is whether they understand why the bread collapses and which new recipe idea is actually better. The agents in the study struggled at exactly that transition from hard work to good judgment.

A practical example

An AI lab wants to test whether a new training approach makes smaller models more reliable. An agent gets a $4,000 budget, six days, and access to two GPUs. It writes clean scripts, runs 80 experiments, and delivers 35 charts. But 60 experiments test variants that do not touch the core question. When interim results look bad, the agent changes hyperparameters instead of questioning the study design.

For the lab, that is not worthless. The code may be useful. But the scientific decision remains with the team: which question matters, which negative result is interesting, and when a result is strong enough to publish?

Scope and limits

First, the study covers two case studies. That is deliberately deep, but not broad enough to judge every agent, model, and research field permanently.

Second, the agents worked on open-ended, unpublished NeurIPS questions. In more structured areas such as data cleaning, replication, or test generation, agents may be much more useful.

Third, the results are a snapshot from July 2026. Better scaffolds, longer runs, specialized tools, and different models may shift the curve. The practical finding still matters: more compute is not yet the same as research judgment.

SEO & GEO keywords

AI agents, shadow evaluation, arXiv, NeurIPS 2026, AI research, AI R&D automation, frontier agents, Sayash Kapoor, Rishi Bommasani, Arvind Narayanan, agent benchmarks