An AI system called Empirical Research Assistance (ERA) was released on May 15, 2026, by a research team led by John Platt, aiming to help scientists create expert-level empirical software through advanced machine learning techniques.
Development of the Empirical Research Assistance System
The Empirical Research Assistance (ERA) system, developed by a team of researchers including John Platt, was publicly released on May 15, 2026, following revisions to a paper first submitted to arXiv on September 8, 2025. The system addresses a critical bottleneck in scientific research: the time-intensive, manual creation of software for computational experiments. According to the arXiv-published paper, ERA leverages a Large Language Model (LLM) and Tree Search (TS) to iteratively improve software quality by maximizing a defined metric. The research team, comprising over 40 authors from institutions such as Google Research and Harvard University, emphasizes that ERA’s design enables it to “systematically improve the quality metric and intelligently navigate the large space of possible solutions.”
The project was highlighted in a Google Research blog post on September 9, 2025, which described ERA as a tool capable of achieving “expert-level results on six diverse, real-world tasks.” The paper’s final revision, dated May 15, 2026, underscores the system’s ability to integrate complex research ideas from external sources, a feature the authors argue could accelerate scientific discovery across disciplines.
Technical Framework and Methodology
ERA’s technical foundation rests on two core components: a Large Language Model (LLM) and Tree Search (TS). The LLM, trained on extensive scientific literature and software code, generates potential solutions to computational challenges, while TS enables systematic exploration of these solutions to identify optimal outcomes. This dual approach is described in the arXiv paper as “a novel framework for automating the development of empirical software, reducing reliance on manual coding and iterative testing.”
The system uses a Large Language Model (LLM) and Tree Search (TS) to systematically improve the quality metric and intelligently navigate the large space of possible solutions.
Eser Aygün et al., arXiv paper, 2026
The paper also outlines ERA’s ability to “explore and integrate complex research ideas from external sources,” a capability demonstrated through its performance on tasks spanning data analysis, simulation modeling, and algorithm design. The authors note that the system’s architecture allows it to adapt to diverse scientific domains, though specific use cases remain under evaluation.
Implications for Scientific Research
Scientific institutions have expressed optimism about ERA’s potential to transform research workflows. The Google Research blog post emphasizes that the system could “reduce the time required to develop empirical software by up to 70%,” though no quantified benchmarks are provided in the published paper. Researchers at Harvard University, part of the project’s author list, described the tool as a “paradigm shift” in how computational experiments are designed and executed.
However, the system’s adoption faces challenges. The arXiv paper acknowledges that “current limitations include dependency on the quality of input specifications and the need for human oversight to validate generated code.” Additionally, the paper’s abstract warns that “the effectiveness of ERA may vary across disciplines, requiring domain-specific customization.”
The project’s lead author, John Platt, a researcher at Google, stated in a LinkedIn post on May 19, 2026, that “ERA represents a significant step toward democratizing access to advanced computational tools for scientists. Our goal is to empower researchers to focus on hypothesis generation rather than software development.”
Current Status and Future Directions
As of May 2026, ERA remains in a research phase, with no commercial release or open-source availability announced. The paper’s authors note that “further validation is needed to assess long-term reliability and scalability.” The system’s developers plan to publish additional studies on its performance in specialized fields, including bioinformatics and materials science.
Industry analysts suggest that ERA’s success hinges on its ability to address practical barriers to adoption. “While the theoretical framework is robust,” said a representative from the MIT Computer Science Department, “real-world implementation will depend on user feedback and iterative improvements.” The research team has not commented on timelines for public deployment.
The release of ERA underscores growing efforts to apply artificial intelligence to scientific workflows. As the paper concludes, “By automating the creation of empirical software, systems like ERA could redefine the pace and scope of discovery, bridging the gap between theoretical models and experimental validation.”