Developer

This pipeline is developed and integrated in BMAP by the author of https://doi.org/10.1002/advs.202417066

Workflow

System architecture. a) Traditional data-driven research paradigm (UNIQUE), where question generation, coding, environment setup, and result evaluation all require human involvement. b) DREAM system based on large language models (LLMs), consisting of multiple functional modules. Blue boxes: LLM-based modules; white box: non-LLM modules; orange outlines: modules using the external tool library BMAP; dashed box: modules containing submodules.

Reference

L. Deng, Y. Wu, Y. Ren, H. Lu, Autonomous Self-Evolving Research on Biomedical Data: The DREAM Paradigm. Adv. Sci. 2025, 12, 2417066. https://doi.org/10.1002/advs.202417066

Background

The exponential growth of biomedical data, coupled with the complexity of analytical tools and disparities in researcher proficiency, has significantly hindered the efficient identification of meaningful scientific questions. While the emergence of Large Language Models (LLMs) has facilitated the development of semi-autonomous research co-pilots across various domains, these existing systems remain heavily reliant on human intervention, lacking capabilities for fully autonomous question generation, advanced iterative reasoning, and automatic environment configuration. To address these limitations and enable large-scale scientific breakthroughs, this study proposes DREAM, a fully autonomous, data-driven, and self-evolving research system capable of continuously interpreting data, generating and refining scientific questions, and executing end-to-end analytical tasks without human intervention.

Result

The evaluation demonstrates that DREAM significantly outperforms existing baselines in both autonomous question generation and iterative self-evolution. While the initial questions generated by the questionRaiser module ranked second only to top-tier peer-reviewed articles—surpassing both GPT-4 and graduate students—the deepQuestioner module further enhanced performance through continuous evolution, ultimately exceeding the difficulty scores of published articles by 5.7% and demonstrating substantial gains in originality. Linear regression analysis across four evolutionary rounds confirmed a significant upward trend in question quality, with the average scores of evolved questions surpassing the benchmark of top-tier articles after just two iterations. These findings collectively validate DREAM’s capacity to generate high-quality, complex scientific questions and its potential to exceed the capabilities of human researchers through iterative self-evolution.

v1.0: Complete the basic process

Usage

1. Select the application 'DREAM'
2. Click on the 'Analysis' button
3. Fill in the request parameters
4. Format your data files according to the default template and compress them into a zip archive
5. Upload the data file