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

• Iteratively formulates scientific hypotheses, design and executes experiments, analyzes and visualizes data

• Progressive agentic tree-search methodology by experiment manager agent
  

• Tree search overcomes in "short-sighted experimentation" -> closer to open-ended hypothesis generation, stepping-stone collection and iterative hypothesis refinement

Experiment Progress Manager

• Coarse-grained experiment management
  *

• Agentic tree search-based exploration

  • With predefined probability a buggy node is chosen.
  • System uses a best-first search strategy when choosing non-buggy node
    

• VLM

  • Provides immediate feedback on generated figures
  • Evaluates figures and the captions

• Stepping stone: Saves intermediate experimental results to serve as building bocks for future rounds

• Agents:

  • Experimenter: writing / executing code for simulations or data analyses
  • Reviewer: "blind" peer review to identify weaknesses
  • VLM: Looks at the figures

• Self correction by code sandboxing: execution / error handling phase: feeds back the traceback to the agent to fix.