What would happen if a community tried to reproduce, experiment by experiment, thousands of studies from a major artificial intelligence conference? That was the question behind the ICML 2026 Open Reproductions challenge, organized by Hugging Face and alphaXiv. The result offers an uncomfortable—but highly useful—look at what it really means to verify science in the age of AI agents.
An experiment to review an entire conference
ICML 2026 received 23,918 papers and accepted 6,352, approximately twice as many as the previous year. This growth is related, at least in part, to tools capable of reading papers, writing code, running experiments, and preparing reports in a short time.
The problem is that review capacity is not growing at the same pace. Most reviewers work voluntarily and do not always have time to check every mathematical proof, experimental setup, or line of code.
A paper can receive a positive evaluation even when the reviewer admits they did not carefully check all of its proofs.
To test this situation, the challenge invited the community to reproduce ICML 2026 papers between July 15 and August 2, 2026. Participants could use coding agents such as Claude Code, Codex, Cursor, or Pi to analyze the papers, implement their methods, and run the tests.
Each participant received $20 in Hugging Face compute credits. In total, 2,962 jobs were launched in the cloud. When a paper depended on private data or unpublished models, teams could create small reproductions with synthetic data that preserved the main properties of the original experiment.
What the reproduction teams found
The challenge brought together 1,221 people and examined 2,200 papers. The results show that reproducibility is not an all-or-nothing category.
- 1,103 papers, or 51%, had at least one independently verified claim.
- 266 papers were fully reproduced, with all extracted claims confirmed.
- 632 papers were partially reproduced and had no refuted claims.
- In total, 3,978 individual claims were confirmed through real experiments.
- 496 papers, or 23%, had at least one refuted or challenged claim.
- In 49 papers, none of the claims could be verified and all were refuted.
- In 242 papers, different teams reached opposing conclusions about the same claims.
The remaining cases fell into an intermediate zone. In 502 papers, there was only evidence at a reduced scale, while in 280, nothing conclusive could be established, mainly because the code, data, or models were unavailable.
This reveals something important: just because a team can run the code does not automatically mean it has confirmed the research. Decisions about scalability, metrics, data, and testing conditions can completely change the result.
Mathematical and experimental errors that did appear
Of the participants, 35 said they had found a refutation. The organizing team reviewed those cases adversarially: it reread the papers, examined work logs, derived the mathematics, and, when necessary, reimplemented the experiments.
One of the most striking cases involved the paper “Towards Optimal Robustness in Learning-Augmented Paging,” presented early in the publication process. The paper claimed a robustness guarantee for its algorithm, but a reproduction observed that an additive term grew with the size of the problem. After expanding the tests to k = 1,024, the organizers confirmed that the growth was not experimental noise, but a sign that a specific step in the proof did not hold.
Another paper, “Attention's Forward Pass and Frank-Wolfe,” claimed that certain token-associated particles collapsed to the origin under specific conditions. Three teams found counterexamples, but at very different points—after 224, 3,800, and 6,416 steps. The shorter checks did not detect the problem.
The finding points to a common trap in AI research: a result can appear correct for thousands of iterations and fail later. If the experiment ends too soon, the wrong conclusion can look convincing.
A difference between theory and code also emerged. In “Self-Distillation Enables Continual Learning,” the theoretical section analyzed the reverse KL divergence, while the published code used the forward KL by default. According to the authors, that was the setting that had generated the paper's results. In addition, the reproduction could not replicate the main four-percentage-point improvement using the available data and code.
In “Do Transformers Need Three Projections?”, one participant discovered that around 66% of the evaluated positions corresponded to padding tokens, EOS. Because these tokens tend to produce a loss close to zero, they artificially reduced perplexity. After correcting the evaluation, the supposed 3.1% quality cost became approximately 9.4%.
AI can also make mistakes when refuting a paper
Reproducing a paper is not simply about trying to prove that it is wrong. Agents can also make mistakes when building a case against it.
In one case, a work log claimed that a paper's method was twice as slow as its baseline. Upon reviewing the analysis, the team found an arithmetic error: it had compared the time for a single trajectory with the time for a batch of 50 trajectories.
After correctly normalizing the measurements, the data itself confirmed the eightfold speedup reported by the authors.
A reproducible result does not depend only on having more compute. It also requires comparing equivalent quantities and reviewing the experiment's assumptions.
The organizers began contacting the authors of each confirmed finding. The initial responses were positive: several authors acknowledged the detected problems, two arXiv corrections were being prepared, and in one case, the author had already quietly fixed the error in a newer version of the paper.
What role is left for humans?
The conclusion is not that agents can completely replace reviewers. Rather, the challenge suggests that human work is changing shape.
The agents were able to read papers, write implementations, launch jobs in the cloud, and produce reports. But they also got stuck in local loops, confused behaviors that depended on scale, and built complete refutations from unit errors.
The most reliable results emerged when a person directed the process: questioning an assumption, changing the direction of the investigation, or deciding that an experiment made no sense before spending a week of compute.
Some evaluations remain difficult to automate. One of the papers studied image generation under extreme quantization. Numerical metrics indicated that the system did not collapse, but the relevant question was visual: were the images still useful to a person?
In that case, the agent created an interface for reviewing the images, and a human personally evaluated 128 pairs. The agent then checked the consistency of the annotations. The machine accelerated the process, but perceptual judgment still depended on a person.
This looks more like the work of a professor or principal investigator than that of a simple operator. The human defines the environment, provides access to data and compute, asks difficult questions, and corrects the course. The agent handles many repetitive tasks at high speed.
A new way to evaluate research
Hugging Face's challenge does not prove that all AI papers are unreliable. It does show that scientific claims need more than a superficial review, especially as the number of publications grows with the help of automated tools.
The most promising combination appears to be agent-assisted review, public logs, executable code, and human supervision. Agents can dramatically expand verification capacity, but they still need direction, judgment, and critical evaluation of their own results.
In other words, AI does not eliminate scientific review. It turns it into a broader, more experimental, and more collaborative process. The question is no longer only whether a paper was accepted, but what happens when someone tries to reproduce each of its claims.