AI Scientist-v2 is an autonomous scientific research system that generates hypotheses, designs experiments, analyzes results, and writes peer-reviewed scientific papers end-to-end using agentic tree search.
The system addresses the challenge of automating the full scientific discovery pipeline by removing dependence on human-authored templates and generalizing across machine learning domains. It employs a progressive agentic tree search guided by an experiment manager agent to explore research directions autonomously. The tool generates research ideas, executes experiments through LLM-written code, and produces complete scientific manuscripts suitable for workshop submission and peer review.
Developers should understand that AI Scientist-v2 prioritizes open-ended exploration over guaranteed success. The README explicitly notes that it does not necessarily produce better papers than its predecessor, especially when strong templates are available. The predecessor achieves higher success rates by following well-defined templates, while this version takes a broader exploratory approach with lower success rates. AI Scientist-v2 suits researchers interested in open-ended scientific discovery rather than those with clear, well-defined objectives and strong foundational templates. The tool requires running on Linux with NVIDIA GPUs using CUDA and PyTorch, and it supports OpenAI, Gemini, and Claude models through various API configurations. A critical consideration is that the system executes LLM-generated code, which poses security risks including potential use of dangerous packages, uncontrolled web access, and unintended process spawning. The README strongly recommends running the tool within a controlled sandbox environment such as a Docker container.
The project maintains active development with structured documentation including a research paper, blog post, and workshop experiment repository. The codebase provides clear installation procedures and comprehensive API key configuration for multiple LLM providers, indicating ongoing maintenance and practical usability considerations.
