BEACON: How Benchmarking Combats the Reproducibility Crisis in Modern Science

I once had the opportunity to have lunch with Nobelist Murray Gell-Mann at a US National Academy of Sciences meeting. Gell-Mann’s many notable breakthroughs are crowned by his discovery of quarks, an idea that helped transform modern physics. I couldn’t help asking Gell-Mann what he felt at the moment he realized the implications of his discovery. His answer was not triumph nor certainty. He said he had been afraid to be wrong.

That sentence has stayed with me. Fear of being wrong should be one of the most important scientific emotions: not fear that paralyzes inquiry, but fear that disciplines it. Fear that forces us to ask: Have I fooled myself? Would this result survive a different dataset, a different laboratory, a different analyst, a different year?

While skepticism of our own research is important, the publishing landscape rewards exciting claims and new discoveries, which can lead to overhype and questionable research practices. Benchmarking can help validate those claims, leading to increased scientific rigor and reproducibility. Our new initiative, Benchmarking, Evaluation, and Assessment Consortium for Science (BEACON), aims to advance this goal by building bridges among siloed benchmarking efforts and enabling validation activities to work synergistically.

Fear of Being Wrong

Much of modern science behaves as if skepticism-fueled-fear were optional. We often continue the enterprise by suspending disbelief. We read papers, cite them, build grants from them, train AI models on their data, and sometimes launch clinical programs based on them, while knowing (or at least we ought to know) that much of the literature may not be reproducible or replicable. Physician-scientist John Ioannidis’s pioneering 2005 essay “Why most published research findings are wrong” did not empirically prove what its title suggests, but it gave a formal warning: Under common conditions of low prior probability, low power, multiple testing, and bias, many statistically significant findings are more likely to be false than true.¹ This should be taken as a warning not to drink our own Kool-Aid simply because a result has appeared in print.

Empirical studies have made the warning harder to ignore. In psychology, the Open Science Collaboration reported that among 100 statistically significant findings from articles published in 2008 across three psychology journals, of which 97 percent had statistically significant results, only 36 percent achieved significance in the same direction upon replication.² In preclinical oncology, Amgen researchers were able to confirm the findings of only 11 percent of 53 landmark cancer papers.³ The Reproducibility Project: Cancer Biology reported a success rate of 46 percent in replicating 112 findings. The rate of replication success differs by field, definition, and methodology, but the broad range is sobering. In many areas such as preclinical research, a conservative estimate is that 50 percent of claims may be inaccurate, inflated, under-specified, or non-replicable.⁴ The precise number matters less than the cultural fact. We know this, yet our publication system still rewards novelty more consistently than reliability. We continue doing science as if the high rate of non-replicability isn’t a serious problem, trusting in the promise that science is self-correcting, while remaining largely oblivious to the effort, money, and opportunity costs that this complacency entails.

When Science Fails to Correct Itself

The problem is not necessarily bad actors, but bad structure. A scientific ecosystem organized around who is first at discovery naturally produces incentives to overclaim. Small datasets, flexible analysis pipelines, selective reporting, reagent variability, insufficient blinding, p-hacking, hypothesizing after the results are known (HARKing), and unpublished negative results (the “file drawer” effect) all push research in the same questionable direction. In machine-learning-based science, the problem that training with non-reproducible data generates non-generalizable models, is compounded by new opportunities for non-replicable models: data leakage, inadequate separation of training and test sets, overfitting, illegitimate features, and test sets that do not represent the target population.⁵

Perhaps the cruelest part is that non-replicable science does not simply disappear—if the findings sound exciting, they gain traction and become world-renowned. A recent study found that papers that failed to replicate in major replication projects were cited more often than papers that did.⁶ The scientific record, then, is not self-correcting at the speed we like to imagine. At times it does the opposite, amplifying the most compelling errors.

Benchmarking was born, in part, as a response to this dilemma.

When I founded the Dialogue for Reverse Engineering Assessment and Methods (DREAM) Challenges in 2006, the motivation was simple. In systems biology, too many papers claimed to have inferred the gene-regulatory networks underlying biological systems. Each method looked impressive in its own paper. Each paper had its own dataset, metric, and validation strategy. But could these algorithms generalize? Could they predict what they had not seen? Could they perform as successfully on a blind test set held back from the developers?

DREAM borrowed consciously from the methodology championed by Critical Assessment of protein Structure Prediction (CASP), which was founded by University of Maryland computational biologist John Moult and colleagues in response to exaggerated claims of success in protein-structure prediction. CASP’s genius was not only that it posed a hard problem, but that it posed the problem fairly. Every two years, participants predicted protein structures that were not yet public; independent assessors then compared predictions against experimental structures.

That framework changed an entire field. For years, CASP allowed protein-structure prediction to take its own pulse. It showed where progress was real, where it was incremental, and where claims exceeded evidence. Then AlphaFold—the Nobel Prize-winning deep learning system to predict 3D protein structures—entered the same arena, under the same norms of blind assessment, and made a leap that could be recognized because the field had built the measurement system to recognize it. AlphaFold2 achieved accuracy competitive with experimental errors, but that statement only has meaning because CASP had created a common evaluative language long before the breakthrough arrived.

DREAM applied the same principle to biomedical modeling: Pose a problem, withhold the answer, and invite the community to solve it. Then, researchers could score the submissions objectively against the known answer and publish the lessons learned. The point was not to crown a winner, but rather to discipline claims. A well-designed challenge can reveal that no method is best in all settings; that ensembles outperform individual models; that a problem is not yet solvable; or that apparent progress was a mirage caused by poor validation. The benefits of this disciplined process are unbiased method assessment and comparison, determining whether a scientific question is solvable, sampling the space of approaches, accelerating research, and building communities around shared data and shared questions.

Benchmarking is a social technology as much as a statistical one. It converts private self-assessment into public, structured comparison. It replaces “trust me” with “test me.” It forces method developers to face the one audience they cannot persuade with rhetoric: unseen data.

This is especially important when using artificial intelligence (AI) in biomedical research. Consider claims that AI could soon replace radiologists.⁷ The Digital Mammography DREAM Challenge offered a more disciplined test.⁸ Algorithms were evaluated on large datasets, with open and reproducible code, and organizers rather than model developers performed the evaluations. In that challenge, no method outperformed radiologists alone, but combinations of AI and radiologist assessment improved performance. The lesson was not anti-AI. It was pro-evidence.

Today, however, domain-specific biomedical benchmarking efforts have proliferated, creating a fragmented landscape. A growing number of initiatives across fields—from protein structure prediction and molecular design to genomics, imaging, and clinical AI—have built valuable communities, datasets, and evaluation practices. The proliferation is a sign of success, but it also dilutes impact. Communities reinvent metrics, governance models, data-sharing practices, leaderboards, evaluation protocols, and publication norms, and they compete for funding and community attention. What should become cumulative progress instead dissolves into a fragmented and siloed effort.

This is why we need a coordinated forum such as BEACON for benchmarking in science. BEACON was conceived by leaders of open science and benchmarking initiatives including Moult (Founder and Chair, CASP), myself (Founder and Chair, DREAM Challenges), Pablo Meyer (Vice Chair, DREAM Challenges), Julio Saez-Rodriguez (Director, DREAM Challenges), Patrick Walters (Chief Scientist, OpenADMET), Aled Edwards (CEO, Structural Genomics Consortium, Founder CACHE), Luca Foschini (President, Sage Bionetworks) and Peng Fu (CEO, Conscience). The goal of BEACON is to align the many benchmarking initiatives that have emerged in response to the reproducibility crisis—including those led by the founding members and others interested in joining the consortium—while respecting their distinct identities, goals, methods, and communities. BEACON is not intended to centralize ownership or impose a single model. Rather, it aims to create a forum where communities with established expertise in rigorous evaluation can share standards, reuse infrastructure, coordinate best practices, and speak with a stronger collective voice. BEACON is structured as a limited partnership managed by a governance committee that includes its founding members. Facilitated by the nonprofit Conscience, in partnership with leading organizations across academia, industry, and open science, BEACON seeks to strengthen the benchmarking ecosystem while preserving the independence and character of each participating initiative.

Benchmarking in the Age of AI

The need for benchmarking is becoming urgent because AI is changing what must be evaluated. We are no longer assessing only algorithms that predict protein folds, gene networks, drug responses, or medical images. We are now dealing with foundation models trained across heterogeneous data modalities including clinical records, imaging, molecular data, and scientific text. We are dealing with AI agents that can plan, retrieve, write code, generate hypotheses, and propose experiments. We are dealing with biomedical tasks where there may be no simple ground truth or gold standard.

How do we benchmark a model that suggests mechanisms of disease? How do we evaluate an AI agent that searches the literature and decides which papers contain non-replicable components? How do we score an AI co-scientist that proposes experiments when no ground truth yet exists or distinguish a useful hypothesis from a persuasive hallucination?

These questions will not be solved by one laboratory or one leaderboard. They require a think tank of community members who have spent years confronting the problem of measuring science: statisticians, experimentalists, challenge organizers, journal editors, funders, regulators, clinicians, data stewards, and AI scientists. BEACON aims to convene such a group to help define when blind prediction is possible, when prospective validation is needed, when human expert adjudication is unavoidable, and when repeated temporal or federated testing is the only plausible way forward.

Science does not need less ambition. It needs more healthy skepticism, or in Gell-Mann’s words, fear of being wrong—not the theatrical humility over-abundant in LinkedIn posts, but operational humility. These include held-out data, independent assessment, transparent code, preregistered protocols where appropriate, community benchmarks, and metrics that survive contact with reality. Gell-Mann’s fear was not a weakness. It was a virtue. The problem with non-replicable science is not that scientists make mistakes. Mistakes are inevitable. The problem is that we have built too many systems that allow mistakes to masquerade as discoveries for too long.

Benchmarking is one way to shorten that interval. It gives doubt an instrument. It gives communities a common ruler. In an era of foundation models and scientific AI agents, it may be one of the most effective tools we have to keep discovery anchored to truth. In a time when scientific claims can spread faster than they can be verified, this discipline is not optional. It is essential.