Search papers, labs, and topics across Lattice.
SciDER is introduced as a data-centric agent system designed to automate the scientific research lifecycle, focusing on processing raw scientific data. It employs specialized agents that collaboratively parse and analyze data, generate hypotheses, design experiments, and execute code. Evaluations on three benchmarks demonstrate SciDER's superior performance in data-driven scientific discovery compared to general-purpose agents, attributed to its self-evolving memory and critic-led feedback loop.
Forget general-purpose agents – SciDER leverages specialized agents, self-evolving memory, and critic feedback to achieve state-of-the-art performance in data-driven scientific discovery.
Automated scientific discovery with large language models is transforming the research lifecycle from ideation to experimentation, yet existing agents struggle to autonomously process raw data collected from scientific experiments. We introduce SciDER, a data-centric end-to-end system that automates the research lifecycle. Unlike traditional frameworks, our specialized agents collaboratively parse and analyze raw scientific data, generate hypotheses and experimental designs grounded in specific data characteristics, and write and execute corresponding code. Evaluation on three benchmarks shows SciDER excels in specialized data-driven scientific discovery and outperforms general-purpose agents and state-of-the-art models through its self-evolving memory and critic-led feedback loop. Distributed as a modular Python package, we also provide easy-to-use PyPI packages with a lightweight web interface to accelerate autonomous, data-driven research and aim to be accessible to all researchers and developers.
