# The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search

**Authors:** Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, David Ha  
**arXiv:** [2504.08066](https://arxiv.org/abs/2504.08066)  
**Date:** 2025-04-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2504.08066v1/

## Key points

- Artificial Intelligence (AI) is revolutionizing scientific research
- The AI Scientist-v2 autonomously generates peer-reviewed workshop papers without human-authored code templates
- Progressive agentic tree-search methodology allows for hypothesis formulation, experiment design, data analysis, and manuscript authoring without human intervention
- Integration of Vision-Language Model (VLM) feedback loop enhances content and aesthetics in figures
- One AI-generated manuscript surpassed the average human acceptance threshold at a peer-reviewed ICLR workshop
- Identified strengths include exploration of temporal consistency regularization and choice of synthetic arithmetic tasks; areas for improvement include clarifying descriptions, addressing omitted references, improving experimental evaluation with real-world tasks and longer sequences, and resolving dataset overlap issues
- Generation process focused on core machine learning topics and real-world applications such as finance, psychology, agriculture, environmental science, and public health
- The AI Scientist-v2 represents a significant advancement in autonomous scientific discovery technologies with open-sourced code to encourage further development

## Summary

In the realm of scientific discovery, Artificial Intelligence (AI) is revolutionizing the way research is conducted. The AI Scientist-v2, an advanced agentic system, has been developed to autonomously generate peer-reviewed workshop papers. This system goes beyond its predecessor by eliminating the need for human-authored code templates and demonstrating effective generalization across various machine learning domains. One key feature of The AI Scientist-v2 is its progressive agentic tree-search methodology managed by a dedicated experiment manager agent. This allows the system to iteratively formulate hypotheses, design experiments, analyze data, and author scientific manuscripts without human intervention. Additionally, the integration of a Vision-Language Model (VLM) feedback loop enhances the AI reviewer component for iterative refinement of content and aesthetics in figures. To evaluate the capabilities of The AI Scientist-v2, three fully autonomous manuscripts were submitted to a peer-reviewed ICLR workshop. Remarkably, one manuscript surpassed the average human acceptance threshold, marking a significant milestone in AI-generated research successfully navigating peer review. This achievement underscores the increasing potential of AI in driving scientific breakthroughs and enhancing research productivity on a large scale. The authors identified strengths and weaknesses in their paper assessment. While they appreciated the exploration of temporal consistency regularization and the choice of synthetic arithmetic tasks for testing hypotheses, areas for improvement were noted. These included clarifying unclear descriptions, addressing omitted references, improving experimental evaluation with real-world tasks and longer sequences, and resolving issues related to dataset overlap. The generation process for the workshop-accepted paper involved idea generation phases focused on core machine learning topics and real-world applications such as finance, psychology,
agriculture,
environmental science,
and public health. Three promising initial ideas were selected based on alignment with the workshop theme and potential interest. The system autonomously executed experimental pipelines using parallelized agentic tree search multiple times to produce high-quality manuscripts for submission. Overall, The AI Scientist-v2 represents a significant advancement in autonomous scientific discovery technologies. By open-sourcing their code to encourage further development in this transformative field, the authors anticipate that AI will continue to play a crucial role in accelerating scientific progress and benefiting society at large.