Meet NovelSeek: a unified multi-agent framework for autonomous scientific research from hypothesis generation to experimental verification

Scientific research across fields such as chemistry, biology and artificial intelligence has long relied on human experts to explore knowledge, generate ideas, design experiments and improve results. However, as the problem becomes more complex and data intensive, discovery becomes slower. Although AI tools such as language models and robotics can handle specific tasks, such as literature search or code analysis, they rarely cover the entire research cycle. Bridging the gap between the generation of ideas and experimental verification remains a key challenge. In order for AI to develop science independently, it must propose hypotheses, design and execute experiments, analyze results, and refinement methods in iterative loops. Without this integration, AI could create disconnected ideas that rely on human supervision verification.

Before introducing a unified system, researchers relied on separate tools at each stage of the process. Large language models can help find relevant scientific papers, but they do not go directly into experimental design or analysis of results. Robotics can assist in automating physical experiments, while coding libraries such as Pytorch can help build models. However, these tools run independently of each other. From forming ideas to verifying them through experiments, no single system can handle the entire process. This leads to bottlenecks where researchers have to manually connect points, slow down progress and leave room for opportunities for errors or missed. The requirements of integrated systems that can handle the entire research cycle can be clearly handled.

Researchers from the NovelSeek team at Shanghai Artificial Intelligence Laboratory have developed Novelseekan AI system designed to automatically run the entire scientific discovery process. NovelSeek consists of four main modules, which are connected in series: a system that generates and perfects the research ideas, a feedback loop where human experts can interact and refine these ideas, a method of turning ideas into code and experimental plans, and a process of conducting multiple rounds of experiments. What makes Novelseek stand out is its versatility. It spans 12 scientific research tasks, including predicting chemical reaction yields, understanding molecular dynamics, predicting time series data, and processing functions such as 2D semantic segmentation and 3D object classification. The team designed NovelSeek to minimize human participation, speed up discovery and deliver consistent high-quality results.

The system behind NovelSeek involves multiple professional agents, each focusing on a specific part of the research workflow. “Investigation Agent” helps the system understand problems by searching for scientific papers and identifying relevant information based on keywords and task definitions. It first conducts extensive investigations of the paper to adjust its search strategy, and then goes deeper by analyzing the full text document for detailed insights. This ensures that the system captures both general trends and specific technical knowledge. Code Review Agent examines existing code bases, whether sourcing from public repositories such as Github, to understand how the current approach works and identify areas of improvement. It checks the structure of the code, looks for errors, and creates summary that helps the system build on past work. “Idea Innovation Agent” produces creative research ideas that drive the systematic exploration of different approaches and refine them by comparing them with relevant research and previous results. The system even includes a “plan and execution agent” that turns ideas into detailed experiments, handles errors during testing and ensures smooth execution of multi-step research plans.

NovelSeek has achieved impressive results in various tasks. In chemical reaction yield forecasts, NovelSeek increased from 24.2% (±4.2 change) at baseline to 34.8% (variance of ±1.1 change of ±1.1), which usually takes several months for human researchers to achieve progress within 12 hours. In the prediction of enhancer activity, a key task in biology, NovelSeek increased the Pearson correlation coefficient from 0.65 to 0.79 in 4 hours. For 2D semantic segmentation, the accuracy increased from 78.8% to 81.0% in just 30 hours. These performance improvements, often achieved in the time normally required, highlight the efficiency of the system. NovelSeek also successfully managed large, complex code bases with multiple files, demonstrating its ability to handle research tasks at the project level, not just in small orphan tests. The team has opened the code to make it available to others, test and contribute to its improvements.

Several key points about novel research include:

• NovelSeek supports 12 research tasks including chemical reaction prediction, molecular dynamics, and 3D object classification.
• The accuracy of reaction yield prediction increased from 24.2% to 34.8% in 12 hours.
• The predicted performance of enhancer activity increased from 0.65 to 0.79 in 4 hours.
• The 2D semantic segmentation accuracy increased from 78.8% to 81.0% in 30 hours.
• NovelSeek includes agents for literature search, code analysis, idea generation and experimental execution.
• The system is open source and can make the scientific field repeatable and collaborative.

In short, NovelSeek demonstrates how combining AI tools into a single system can speed up scientific discovery and reduce reliance on human efforts. It connects key steps together, generates ideas, translates them into methods, and tests them experimentally into a simplified process. A work that researchers who have spent months or years could do in days or even hours. By linking each phase of the research into a continuous loop, NovelSeek can help the team move from rough ideas to real-world results. The system emphasizes the power of AI not only to help, but to drive scientific research in a way that can reshape discoveries in many fields.

View paper and GitHub pages . All credits for this study are to the researchers on the project. Also, please stay tuned for us twitter And don’t forget to join us 95k+ ml reddit And subscribe Our newsletter.

Nikhil is an intern consultant at Marktechpost. He is studying for a comprehensive material degree in integrated materials at the Haragpur Indian Technical College. Nikhil is an AI/ML enthusiast and has been studying applications in fields such as biomaterials and biomedical sciences. He has a strong background in materials science, and he is exploring new advancements and creating opportunities for contribution.