At the core of ERA’s groundbreaking capabilities is its fusion of advanced large language models with sophisticated search strategies, enabling it to iteratively refine and optimize scientific code far beyond conventional human capacities. By integrating Google’s Gemini language model with a specialized tree search algorithm—similar to those used in landmark AI systems like AlphaGo—ERA dynamically explores a vast landscape of programming variations. This process allows it to evaluate and select promising code modifications with the explicit purpose of enhancing performance metrics, or “scores,” pertinent to various scientific challenges such as epidemiological forecasts and molecular modeling of proteins.

The system fundamentally operates by starting with an initial baseline of empirical software tailored to a specific research problem. ERA then autonomously proposes and tests numerous incremental modifications, ranging from algorithmic substitutions to the inclusion of novel computational modules. This approach mirrors an accelerated scientific experimentation cycle, where hypotheses are quickly implemented and evaluated based on their quantitative impact on predefined performance criteria. Crucially, ERA incorporates a feedback loop that directs its search towards avenues most likely to yield substantive improvements, an innovation that significantly compresses the timeline for software refinement from months or years to mere hours or days.

What truly distinguishes ERA is its capacity to synergize AI-driven code generation with human scientific insight. Users can input foundational knowledge from academic papers, textbooks, or domain-specific heuristics, which ERA then assimilates into its evolutionary programming process. This fusion of human-guided theory and AI-powered exploration enables the system to uncover subtle, “needle-in-a-haystack” solutions that might elude even the most seasoned researchers due to their complexity or non-intuitive nature.

The practical implications of ERA were demonstrated emphatically through a series of rigorous experiments. In a neuroscientific inquiry, the system was tasked with modeling the activity of over 70,000 neurons within a zebrafish brain, leveraging an existing neuron-modeling toolkit. While a human would face a steep learning curve requiring extensive manual coding and simulation tuning, ERA autonomously assembled and calibrated neural models with high physical fidelity, dramatically accelerating the research cycle. This accomplishment underscores ERA’s potential to transform computational neuroscience by enabling rapid hypothesis testing and model generation at scales previously unattainable.

Further validation came from epidemiological modeling, where ERA produced an ensemble of fourteen predictive models for COVID-19 hospitalizations, outperforming even the established frameworks used by the U.S. Centers for Disease Control and Prevention. This highlights ERA’s profound capacity to process real-world, dynamic datasets and optimize predictive algorithms under the pressing conditions of a global health crisis. Additionally, ERA outstripped existing human-designed methodologies in integrating single-cell RNA sequencing data, revealing new computational strategies that advance the rapidly evolving field of genomics.

From a methodological perspective, the project showcases the power of integrating machine learning paradigms with empirical science workflows. ERA’s architecture achieves a dual purpose: it acts as a tireless software developer, tirelessly iterating over potential designs, and as an autonomous researcher, intelligently navigating the trade-offs inherent in complex scientific programming tasks. This synergy not only enhances computational throughput but also encourages a paradigm shift in how empirical problems are approached, inviting a closer collaboration between AI and domain experts.

The implications of ERA’s automation extend beyond mere efficiency gains. By drastically shortening the time required to explore and test diverse computational ideas, the system liberates scientists to focus on more creative and conceptual challenges—defining fundamental research questions, uncovering novel hypotheses, and addressing pressing societal issues through targeted scientific inquiry. ERA thus acts as an accelerator for the entire scientific ecosystem, increasing the pace of innovation in fields ranging from climate science and epidemiology to materials science and computational biology.

Harvard’s Michael Brenner, a key figure in this initiative, emphasizes that ERA’s most exciting potential lies in its ability to “integrate and recombine research ideas,” pushing scientific software development into territories previously inaccessible through manual experimentation. By harnessing this advanced AI, researchers can pursue bold, high-risk computational strategies that may have been deemed impractical or prohibitively complex. This newfound agility could reshape scientific priorities and expand the frontier of knowledge.

Beyond its immediate research applications, the ERA system exemplifies a broader movement toward the convergence of academia and industry. Through the Catalyst Professorship, Harvard has facilitated Michael Brenner’s collaboration with Google, demonstrating how academic-industry partnerships can yield transformative breakthroughs. This model of shared expertise, resources, and perspectives is becoming increasingly vital in addressing complex multidimensional problems where computational prowess is paramount.

Importantly, ERA’s design approach can be tailored to diverse domains, making it highly adaptable. Whether tackling protein folding dynamics, forecasting environmental phenomena, or optimizing algorithms for autonomous systems, ERA serves as a flexible platform capable of embracing a wide array of empirical research questions. Future iterations may integrate even more advanced theoretical knowledge and domain-specific constraints, further refining its problem-solving capabilities.

In summary, the development of Empirical Research Assistance represents a monumental stride in accelerating scientific progress through AI. By automating the generation and refinement of empirical software, ERA alleviates a chronic bottleneck in the research process, enabling scientists to conduct more exhaustive explorations faster and with greater precision. As this technology matures, it promises to redefine the landscape of computational science and empower researchers worldwide to achieve breakthroughs that were once out of reach.

Subject of Research: Not applicable
Article Title: An AI system to help scientists write expert-level empirical software
News Publication Date: 19-May-2026
Web References:

• Google Research Blog: https://research.google/blog/accelerating-scientific-discovery-with-ai-powered-empirical-software/
• Nature DOI: http://dx.doi.org/10.1038/s41586-026-10658-6
  References:
  Brenner, M. et al. (2026). An AI system to help scientists write expert-level empirical software. Nature.
  Image Credits: Not specified

Keywords
Artificial intelligence, Computer architecture, Computer modeling, Computers, Applied mathematics, Computer science, Software, Supercomputing, Artificial neural networks, Generative AI, Machine learning, Deep learning, Neural net processing

High-temperature superconductors, particularly cuprates, represent a frontier in condensed matter physics, characterized by their unique electronic properties that challenge existing theoretical frameworks. The sheer volume and sophistication of research articles in this field render traditional literature review an arduous endeavor even for seasoned specialists. To evaluate whether LLMs can bridge this complexity, the research team designed a rigorous testing ground, melding computational prowess with expert human insight. This interdisciplinary approach zeroed in on a curated collection of 1,726 carefully selected scientific papers detailing decades of cuprate superconductivity research.

Central to the study was the deployment of six state-of-the-art LLM systems, including notable models such as ChatGPT-4 and Claude 3.5, alongside specialized tools like Google’s NotebookLM and a customized retrieval-augmented generation (RAG) system. These models were tasked with navigating the dense literature, synthesizing nuanced scientific insights, and answering 67 meticulously crafted questions that probed the depths of theoretical understanding in the domain. The questions were developed by a cohort of domain experts to truly test the models’ grasp on critical scientific nuances and theoretical subtleties.

The study unveiled a noticeable performance gradient among the models. Systems such as NotebookLM and the bespoke RAG framework, which operated on databases of vetted scientific papers rather than unfiltered internet data, demonstrated superior capability in extracting and reasoning over precise scientific information. This finding substantiates the importance of anchoring LLMs to reliable, authoritative data sources to curtail hallucinations and unwarranted claims often observed in language models. Despite the impressive textual retrieval capabilities, the models consistently fell short in interpreting complex visual data, such as plots and figures that are foundational to scientific argumentation in physics.

Eun-Ah Kim, the lead Cornell physicist spearheading this inquiry, emphasized that this research underscores the current limitations of LLMs relative to the ambitious goals of artificial general intelligence (AGI). While these models exhibit remarkable abilities in language generation and comprehension, their inability to synthesize multifaceted scientific problems with fidelity signals a key deficiency in artificial cognition. Particularly, the models struggle with integrating diverse conceptual angles and reflecting the inherent complexities and contradictions endemic to frontier scientific problems like those in cuprate superconductivity.

The evaluation also shed light on subtle behaviors detrimental to scientific rigor, including occasional fabrication of bibliographic references and unreliable attribution of claims within LLM outputs. These issues prompt caution in relying exclusively on language models for expert-level scientific inquiry and highlight the necessity of continued human oversight. Enhancing models’ accuracy in reference generation and multimodal understanding, especially in interpreting graphical data, stands as a critical frontier for future AI development.

The research team’s methodology went beyond superficial testing, leveraging a sophisticated database that reflects the historical evolution and contemporary debates in high-temperature superconductivity research. The 67 probing questions incorporated aspects ranging from theoretical modeling to experimental results interpretation, demanding not only recall but also analytical reasoning. This thoroughness ensures the study’s conclusions carry weight in assessing the true potential and current shortcomings of LLM-driven scientific analysis.

Notably, Google’s NotebookLM and the specially designed RAG system leveraged curated datasets to substantially outperform general-purpose LLMs. NotebookLM’s architecture, which allows user-driven document uploads for targeted inquiry, demonstrates a promising paradigm where domain-specific datasets empower AI to act more reliably as a research assistant. This approach of grounding models in trusted knowledge repositories mitigates risks associated with unverified internet-sourced content that can lead to misinformation and superficial answers.

Despite these advancements, the research reveals a critical bottleneck in the AI’s ability to grasp and engage with complex scientific visualizations, an essential ingredient for a comprehensive understanding of high-Tc superconductivity phenomena. Data visualizations encapsulate multidimensional correlations, subtle experimental nuances, and theoretical predictions—factors that remain elusive to existing LLM frameworks. The struggle with visual data highlights a dimensionality gap in AI understanding which current text-centric models have yet to bridge.

The experiment serves as a vital touchstone in framing expectations around LLM applications in scientific disciplines. While the interrogated models display remarkable facility in the synthesis of textual scientific information, their deficiencies remind the research community that AI, in its present form, is a tool best used in concert with human expertise rather than as an autonomous oracle. This perspective aligns with the broader vision that AGI, capable of generalized understanding and reasoning, remains on the horizon rather than an imminent reality.

Looking ahead, the Cornell-led research group, as part of the National Science Foundation’s AI-Materials Institute, aims to leverage these insights to inform the iterative design of more robust and scientifically adept AI systems. By emphasizing multimodal processing capabilities, improved factual grounding, and sophisticated reasoning architectures, the next generation of AI can inch closer to true scientific comprehension. This study thus not only maps the terrain of current AI capabilities but also charts a strategic course for future innovation.

The findings have broader implications beyond the niche of cuprate superconductivity, offering a prototype for evaluating LLMs across other highly specialized scientific domains. As AI permeates academic research, the imperative to quantitatively and qualitatively benchmark its performance grows stronger. The meticulous human expert panel employed in this study offers a replicable model for assessing AI’s adherence to scientific rigor, an essential benchmark for success in the age of automated knowledge extraction.

In conclusion, the Cornell and Google collaborative study presents a nuanced portrait of contemporary LLM capabilities, revealing both potent strengths and notable deficiencies in their scientific world modeling. Their work invites a tempered optimism—acknowledging significant progress while cautioning against overestimation of AI’s current understanding. By anchoring AI advancements in rigorous experimental tests and curated datasets, the scientific community moves closer to harnessing the full potential of these technologies without compromising the integrity of scholarly inquiry.

Subject of Research: Evaluation of large language models’ ability to comprehend specialized scientific literature in high-temperature cuprate superconductivity.

Article Title: Expert Evaluation of LLM World Models: A High-Tc Superconductivity Case Study

News Publication Date: Not explicitly provided

Web References:
Proceedings of the National Academy of Sciences DOI
Cornell Chronicle Story
National Science Foundation AI-Materials Institute

References:
Guo, H. et al. (2026). Expert Evaluation of LLM World Models: A High-Tc Superconductivity Case Study. Proceedings of the National Academy of Sciences. DOI: 10.1073/pnas.2533676123

Keywords

Artificial intelligence, large language models, high-temperature superconductivity, cuprates, scientific literature, AI evaluation, retrieval-augmented generation, NotebookLM, multimodal AI understanding, interpretability, National Science Foundation, AI-Materials Institute

A pioneering step toward overcoming these hurdles is eNRRCrew, a cutting-edge multi-agent artificial intelligence (AI) framework developed through a collaborative effort led by Professors Zhen Zhou and Xu Zhang from Nankai University and Zhengzhou University, respectively. Published in the esteemed National Science Review, eNRRCrew represents a marriage of large language models (LLMs) and machine learning techniques engineered to automate the entire catalyst research workflow. By integrating data mining, predictive analytics, and knowledge synthesis into a unified system, it fundamentally transforms how researchers approach catalyst discovery.

The real power of eNRRCrew lies in its beginning: analyzing an extensive corpus of 2,321 scientific publications related to nitrogen reduction catalysts. This mammoth task, which would otherwise require months of meticulous work from human experts, was fully automated using the AI crew. The resultant comprehensive database catalogs catalyst compositions, structural features, reaction environments, and performance metrics. This vast, structured knowledge base enables the system not only to retrieve pertinent experimental data swiftly but also to train advanced machine learning models capable of forecasting catalyst behavior with unprecedented accuracy.

Delving into the machine learning insights reveals that two physical parameters—the symmetry of the catalyst’s crystal structure, quantified by space group numbers, and the electronegativity differences among constituent elements—are pivotal determinants of catalytic efficiency. These factors extend the frontier of understanding beyond mere empirical observation, offering a rational framework for designing new catalysts anchored in physicochemical principles rather than heuristic guesswork. This breakthrough epitomizes how AI can distill complex phenomena into actionable design criteria, propelling experimental efforts toward higher success rates.

Structurally, eNRRCrew operates through five collaboratively functioning AI agents, each specializing in unique aspects of the research process. An orchestrator oversees task delegation and integration, predictor agents assess potential catalyst yields and efficiencies, a knowledge graph retriever navigates the interconnected scientific data landscape to provide contextual insights and recommendations, and a file handler interprets raw data to generate informative analyses. Researchers communicate with this sophisticated ensemble in natural language, enabling intuitive interactions that transcend traditional database queries or static computational models.

The user experience of eNRRCrew is dynamic and interactive. Scientists can pose complex technical questions, simulate catalyst performance, and request novel candidates tailored to specific reaction parameters. The system’s repertoire includes generating customized visualizations such as data plots, crafting evidence-based summaries that cite primary literature, and autonomously proposing promising new catalysts. Notably, eNRRCrew excelled as a discovery engine by recommending 13 novel catalyst compositions predicted to possess high ammonia yield and operational stability, a feat that surpasses conventional manual exploratory methods.

Among its recommendations, the Mo–W dimer anchored on a Ti₂NO₂ MXene substrate stood out for its predicted stability and catalytic potential—a finding corroborated by sophisticated computational simulations rooted in quantum mechanical principles. These simulations elucidate the interaction mechanisms between the catalyst surface and nitrogen molecules, further validating the AI’s predictive capability. Another novel catalyst, MoFeNC, has successfully transitioned from AI-generated hypothesis to experimental synthesis and validation within the research team’s laboratories, showcasing the framework’s efficacy in bridging theoretical design with practical application.

The multi-agent architecture fundamental to eNRRCrew confers distinct advantages over isolated AI models or traditional experimental paradigms. It embodies a paradigm shift where autonomy, specialization, and collaboration converge, permitting the efficient handling of large data volumes and the synthesis of multidimensional knowledge. According to Prof. Zhou, this approach transcends the limitations of single LLMs by providing modularity and coordinated action, ultimately expediting scientific discovery cycles while ensuring robustness and interpretability.

Beyond ammonia synthesis, the modular design philosophy underpinning eNRRCrew displays remarkable versatility, having already been adapted for a variety of electrocatalytic reactions central to energy conversion and storage technologies. This adaptability underscores its potential to revolutionize multiple domains within catalysis research and materials science, advancing the implementation of sustainable chemical processes at broad scales.

The advent of intelligent, collaborative AI frameworks such as eNRRCrew heralds a new era in scientific inquiry where human creativity and machine precision synergistically propel exploration. By autonomously parsing vast scientific landscapes, generating data-driven hypotheses, and guiding experimental validation, such systems promise to alleviate bottlenecks inherent in discovery-driven disciplines. eNRRCrew exemplifies this vision, accelerating the path toward greener ammonia production and illuminating a blueprint for future AI-enabled research infrastructure.

As global demands for sustainable agriculture and green chemicals intensify, innovations like eNRRCrew provide critical leverage in transforming foundational industrial processes to be more energy-efficient, environmentally benign, and economically viable. This breakthrough not only advances ammonia catalysis but also pioneers an AI-driven methodology that could redefine how complex scientific challenges across disciplines are addressed in the 21st century.

Subject of Research:
Electrocatalyst discovery for sustainable ammonia synthesis via electrochemical nitrogen reduction reaction using multi-agent AI systems.

Article Title:
eNRRCrew: A Multi-Agent AI Framework for Catalyst Discovery in Sustainable Ammonia Production.

Web References:
10.1093/nsr/nwaf372

Image Credits:
©Science China Press

Keywords

Ammonia synthesis, electrochemical nitrogen reduction reaction, catalyst discovery, multi-agent AI, machine learning, large language models, sustainable chemistry, electrocatalysis, computational simulation, crystal structure, electronegativity, MXene, green fertilizers