In the relentless pursuit of scientific discovery, one persistent barrier has stood the test of time: the painstaking manual creation of software that underpins computational experiments. This bottleneck not only slows progress but also demands considerable expertise and time from researchers, diverting their focus from core scientific questions. Addressing this formidable challenge head-on, a groundbreaking advancement has emerged in the form of Empirical Research Assistance (ERA), an artificial intelligence system designed to autonomously generate expert-level scientific software. ERA’s mission transcends simple automation; it strives to maximize the quality metric of the software it creates, thereby accelerating the pace of scientific innovation.
At the heart of ERA’s powerful capabilities lies the synergistic combination of Large Language Models (LLMs) and sophisticated Tree Search algorithms. These two critical components together empower the system to methodically explore vast repertoires of potential software solutions. The use of LLMs allows ERA to comprehend and synthesize complex scientific texts and code snippets, effectively mimicking expert human reasoning. Meanwhile, the Tree Search algorithm orchestrates a strategic navigation through the labyrinthine solution landscape, pruning suboptimal paths while promoting exploration of promising avenues. This meticulous optimization process enables ERA to iteratively refine its software outputs, continually inching closer to designs that excel by domain-specific criteria.
ERA’s effectiveness is not a mere theoretical promise but a demonstrated reality across multiple scientific disciplines. Particularly compelling are its achievements in bioinformatics, where singular-cell data analysis poses immense computational challenges. In this domain, ERA autonomously devised forty novel analytical methods, each surpassing existing human-crafted counterparts. The system’s creations outperformed the highest-ranking methods on a widely recognized public leaderboard, reflecting both innovation and practical superiority. This landmark accomplishment underscores ERA’s capacity to not only replicate expert thought but to transcend it, forging new paths in data interpretation that advance biological understanding.
The impact of ERA extends into epidemiology, a field where predictive modeling is vital for public health preparedness. During the COVID-19 pandemic, forecasting hospitalizations accurately was paramount for resource allocation and policy decisions. ERA generated fourteen innovative epidemiological models, each outclassing the ensemble forecasts produced by the Centers for Disease Control and Prevention (CDC) as well as all individual prediction models submitted for comparison. These results highlight how ERA’s automated approach can enhance crisis response by delivering more accurate and timely predictions, potentially saving lives and optimizing healthcare delivery.
Beyond these domains, ERA demonstrates versatility by producing software that reaches expert-level standards in geospatial analysis—an area critical for environmental science and urban planning—as well as in predicting neural activity in zebrafish, a model organism widely used in neuroscience. Additionally, ERA succeeded in developing novel numerical solutions for integral calculus problems, an essential task in many branches of physics and engineering. Remarkably, it also engineered an innovative rule-based method specifically tailored for time series forecasting, illustrating its ability to generate fully original algorithmic strategies.
The core innovation driving ERA is its ability to integrate complex, external research ideas and methods into its software generation process. This integration is no trivial feat, as it entails digesting diverse scientific inputs, reconciling different analytical frameworks, and synthesizing new, coherent solutions. The system’s proficiency in doing so essentially mimics the intellectual labor of human scientists, but with far greater speed and breadth. Such capability promises to transform the way scientific software is created, shifting from bespoke, manual coding to automated, scalable generation that continuously refines itself toward optimal performance.
ERA’s tree search component deserves particular emphasis for its critical role in making this revolution possible. Tree search algorithms are designed to systematically explore decision trees, where each branch represents a potential step or choice in software design. By evaluating intermediate results against the quality metric, ERA avoids less promising directions, concentrating computational resources on pathways more likely to yield breakthroughs. This mechanism enables the system not only to navigate but also to conquer the immense combinatorial complexity of scientific software design, which traditionally stymies human efforts.
The research team behind ERA demonstrated the system’s wide-ranging utility through rigorous benchmarking across heterogeneous tasks. The benchmarks included not only prediction accuracy and model robustness but also software efficiency and adaptability, underscoring ERA’s comprehensive expertise. Such evaluations confirm that ERA’s approach generalizes well across multiple scientific problems, a crucial attribute for any tool aiming to support diverse research endeavors. This generality sets ERA apart from prior AI-assisted coding tools that are often narrowly specialized.
Crucially, ERA embodies a paradigm shift in the scientific process itself. It transforms the production of scientific software from a limiting step into an accelerative technology, vastly expanding the capacity of researchers to test hypotheses and generate knowledge. By automating the generation of high-quality empirical software, ERA liberates scientists to devote their creativity and intellectual energy to theory development, experimental design, and interpretation, instead of programming logistics. In doing so, ERA not only enhances productivity but may redefine the fabric of scientific discovery.
Beyond the immediate technical achievements, ERA invites profound reflection on the evolving role of artificial intelligence in science. As an autonomous creator of expert-level software, ERA bridges the gap between human conceptual insight and machine precision, suggesting a future where AI systems act as collaborators rather than mere tools. The system’s success raises tantalizing prospects for co-discovery, where human intuition and machine optimization synergize to unlock knowledge inaccessible to either alone. This vision marks a compelling frontier in the integration of human and artificial intelligence.
The introduction of ERA also has implications for scientific education and training. By producing exemplar software designs and novel methodologies, ERA can serve as a didactic resource, exposing researchers and students to cutting-edge algorithmic practices and experimental strategies. This educational dimension amplifies ERA’s impact, fostering a community that benefits from accelerated learning alongside accelerated discovery. As such, ERA may catalyze a virtuous cycle in which enhanced knowledge fuels further innovation in both AI and domain sciences.
Looking ahead, the potential applications of ERA span an extraordinary range of scientific fields. Any domain reliant on computational experiments—from climate science and genomics to robotics and economics—stands to gain from the system’s capabilities. The adaptability and extendibility of ERA’s architecture suggest that it could readily incorporate advances in AI research, such as improved language models or optimization techniques, further sharpening its effectiveness. In this light, ERA emerges not just as a milestone but as a foundational platform for a new era of AI-powered science.
In summary, Empirical Research Assistance represents a transformative breakthrough in the automation of scientific software creation. By harnessing the complementary strengths of large language models and tree search algorithms, ERA navigates and optimizes complex solution spaces to produce high-quality, expert-level empirical software across numerous scientific disciplines. Its demonstrated successes—from bioinformatics to epidemiology and beyond—highlight its revolutionary potential to accelerate scientific progress. As ERA continues to evolve, it promises to reconfigure the landscape of research, empowering scientists worldwide to solve ever more challenging problems with unprecedented speed and sophistication.
Subject of Research: Automated scientific software generation, AI-assisted empirical research, AI in computational science
Article Title: An AI system to help scientists write expert-level empirical software
Article References:
Aygün, E., Belyaeva, A., Comanici, G. et al. An AI system to help scientists write expert-level empirical software. Nature (2026). https://doi.org/10.1038/s41586-026-10658-6
Image Credits: AI Generated
