Idea Generation

Initial Setup

Start with a provided code template that includes a lightweight baseline training run from a popular model or benchmark.

The template also contains LaTeX files with style files, section headers, and basic plotting scripts

Brainstorming

The AI Scientist “brainstorms” a set of novel research directions by leveraging inspiration from evolutionary computation and open-endedness research

Idea Refinement

Multiple rounds of chain-of-thought reasoning and self-reflection are used to refine and develop each idea

Filtering

Ideas are filtered through a Semantic Scholar API search and web access to discard those too similar to existing literature

Final Output

The generated ideas include a description, experiment execution plan, and self-assessed scores for interestingness, novelty, and feasibility