A new framework automates the research and development process by structuring problem selection and experimental design. The approach uses a two-axis model to evaluate projects by Interest and Feasibility, guiding researchers toward a productive Pareto front. This method counters the common pitfall of choosing overly ambitious or trivial questions.

The system frames experimentation as a Bayesian optimization problem, where each experiment aims to reduce uncertainty. Instead of simple trial and error, it models the search space and selects tests that most effectively update belief in competing hypotheses. This algorithmic perspective treats the research cycle as a series of informed decisions.

The AI Scientist-v2 agent implements this philosophy. It generates initial research ideas based on domain constraints and resource limits, then manages a four-stage workflow: preliminary investigation, hyperparameter tuning, main execution, and ablation studies. This structured tree-based experimentation aims to make autonomous research more systematic and less reliant on serendipity.