The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search: Fully Automated Scientific Paper Generation System Passes Peer Review for the First Time
In April 2025, the Sakana AI team released AI Scientist-v2, an end-to-end agentic system capable of autonomously completing the entire process of hypothesis generation, experimental design, data analysis, and paper writing. The system submitted three fully AI-generated papers to the ICLR 2025 workshop, one of which scored above the human average acceptance threshold, becoming the first fully AI-generated paper to pass peer review. Compared to v1, v2 no longer relies on human-written code templates and adopts a progressive agentic tree search method coordinated by a dedicated experiment manager agent.
AI Scientist-v2 marks a critical leap for AI in scientific discovery, from an assistive tool to an autonomous researcher. It demonstrates for the first time that AI-generated papers can meet human-acceptable academic standards, which will profoundly impact research productivity, academic publishing, and knowledge production. The system achieves full automation from hypothesis to paper through agentic tree search and visual language model feedback loops, and the code is open-sourced.
The core innovation of AI Scientist-v2 lies in progressive agentic tree-search, where an experiment manager agent coordinates multiple sub-agents to explore the experimental space in parallel. The system comprises four main modules: hypothesis generator, experiment designer, data analyzer, and paper writer. Compared to v1, v2 removes the dependency on human-written code templates, achieving cross-domain generalization through dynamic code generation and debugging. Additionally, the AI reviewer component integrates a visual language model (VLM) feedback loop to iteratively optimize figure content and aesthetics. Evaluation uses the real ICLR workshop peer review process, not human simulation.
This achievement will profoundly transform the academic publishing and research service industries. For research institutions, AI Scientist-v2 can significantly reduce the labor cost of paper writing and experimental design, accelerating knowledge output. For publishers, it necessitates redefining authorship and review standards. For tech companies, the technology can be integrated into research collaboration platforms, offering one-stop services from data to paper. Meanwhile, the open-source code will drive community-driven development of research automation tools.
It is recommended that research service companies and academic publishers immediately evaluate the integration potential of AI Scientist-v2. A research automation SaaS platform based on this system could be developed, offering paper generation and experiment optimization services to universities and research institutes. Investment directions include: collaborating with Sakana AI to develop industry-specific versions, or building vertical-domain (e.g., drug discovery) research automation tools based on the open-source code.
Attention should be paid to AI Scientist-v2's generalization ability across more disciplines (e.g., biology, chemistry), as well as the reproducibility and novelty of its generated papers. Key signals include: whether the system can publish papers at top-tier conferences (not workshops), changes in acceptance rates by human reviewers, and ethical discussions around AI authorship in academia. Additionally, the system's computational cost and energy consumption are important considerations for practical deployment.