The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search
github.com/sakanaai/ai-scientist-v2- Category
- AI Agents
- Rank
- No. 442Tools index
- Type
- APP
- Builder
- sakanaai
- GitHub
- 7.1k stars
- Date
About
The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search 📚 [Paper] | 📝 [Blog Post]
What it does
It turns a research topic into candidate hypotheses, explores experiments through managed tree search, analyzes resulting data, builds plots, writes a manuscript, and reviews the finished PDF. Separate language and vision model stages handle text, figures, citations, and critique.
Why it's ranked here
This is a substantial research automation pipeline, not a thin prompt wrapper. It connects ideation, parallel experimental exploration, plot generation, citation gathering, manuscript compilation, and review. The verdict is mixed because the project explicitly reports lower success rates than its template-driven predecessor and warns that runs may fail to produce a PDF or review.
What's good
The workflow preserves useful intermediate artifacts: structured ideas, experiment logs, a searchable tree visualization, plots, token accounting, manuscript PDFs, and text and figure reviews. Search controls expose worker count, exploration depth, debugging behavior, and independent drafts. Users can also assign different models to experimentation, plotting, writing, citations, and review.
Tradeoffs
The system executes model-written code, with risks including dangerous packages, uncontrolled web access, and unintended processes. It expects Linux, NVIDIA GPUs, CUDA, PyTorch, PDF tooling, LaTeX tooling, API credentials, and several hours for a full example run. Semantic Scholar works without a key, but rate limits or weaker novelty checking may follow. Open-ended exploration also has lower success than template-based work.
How to use it well
Use it for machine-learning researchers testing open-ended ideas where broad exploration matters more than predictable completion. Start with a tightly scoped topic description, inspect generated hypotheses, tune search breadth and debugging limits, then run inside a controlled sandbox and audit every artifact. It does not replace safe infrastructure, expert scientific judgment, or a guaranteed manuscript pipeline.
Technical notes+
README.md defines a two-stage CLI workflow and documents Linux, CUDA, PyTorch, model credentials, Semantic Scholar integration, tree-search controls, and generated artifacts. launch_scientist_bfts.py orchestrates idea loading, per-run directories, configuration rewriting, best-first tree search, result copying, plot aggregation, token tracking, citation gathering, write-up retries, PDF review, vision review, and child-process cleanup. ai_scientist/llm.py routes requests across OpenAI-compatible, Anthropic, Bedrock, Vertex AI, Gemini, DeepSeek, and Ollama model names with retry and token tracking. ai_scientist/vlm.py base64-encodes images and sends multimodal requests. ai_scientist/perform_plotting.py writes and executes model-generated plotting code, then iterates through reflection rounds. ai_scientist/perform_writeup.py gathers Semantic Scholar citations and compiles LaTeX. ai_scientist/perform_llm_review.py and ai_scientist/perform_vlm_review.py separately assess manuscript text and rendered figures.
Observed
- Primary language
- Python
- Interface
- Command-line scripts for idea generation and the full experiment-to-paper pipeline
- Installation surface
- Conda environment with Python 3.11, CUDA-enabled PyTorch, Poppler, chktex, and pip requirements
- Platform support
- Designed for Linux with NVIDIA GPUs and CUDA
- Model access
- Supports hosted model APIs and local Ollama-backed models
- Output artifacts
- Structured idea JSON, experiment logs, an HTML tree visualization, plots, manuscript PDF, and review files
Read from README.md, requirements.txt, launch_scientist_bfts.py, ai_scientist/llm.py, ai_scientist/vlm.py, ai_scientist/perform_writeup.py, ai_scientist/perform_plotting.py, ai_scientist/perform_llm_review.py, ai_scientist/perform_vlm_review.py.
What it can do
Generate scientific hypotheses autonomously
Research domain or topic area → Scientific hypotheses for investigation
Design and execute machine learning experiments
Research hypothesis and domain specifications → Experimental results and data
Analyze experimental data automatically
Raw experimental data and results → Statistical analysis and insights
Write complete scientific manuscripts
Experimental results and analysis → Formatted scientific papers with sections, figures, and citations
Perform agentic tree search for research exploration
Research objectives and constraints → Optimized research paths and experimental strategies
Execute LLM-generated code for experiments
AI-written experimental code → Code execution results and outputs
Manage multi-agent research workflows
Research tasks and agent configurations → Coordinated research activities and results
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.