GitHub - SakanaAI/AI-Scientist-v2: The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search
5 hours ago
- #Agentic Tree Search
- #Autonomous Scientific Discovery
- #AI Research Automation
- The AI Scientist-v2 is a fully autonomous end-to-end agentic system capable of generating hypotheses, running experiments, analyzing data, and writing scientific manuscripts, with the first AI-written workshop paper accepted through peer review.
- Unlike its predecessor v1, which relies on human-authored templates, v2 generalizes across ML domains, uses a progressive agentic tree search guided by an experiment manager agent, and is designed for open-ended scientific exploration, though it may have lower success rates.
- The system requires a sandboxed environment (e.g., Docker) due to risks like executing LLM-written code, using dangerous packages, and uncontrolled web access, with installation on Linux using NVIDIA GPUs, CUDA, and PyTorch.
- Setup involves creating a conda environment, installing dependencies, and configuring API keys for models (e.g., OpenAI, Gemini, Claude via AWS Bedrock) and optional Semantic Scholar for literature search and novelty checking.
- Usage includes generating research ideas via a Markdown topic description and running the main pipeline with agentic tree search, where parameters like num_workers and steps are configured in bfts_config.yaml, leading to PDF output.
- Estimated costs are around $15–$20 per run for experimentation with Claude 3.5 Sonnet and an additional $5 for writing, with citation provided for research use and mandatory disclosure of AI involvement in manuscripts.