GEAR: Genetic AutoResearch for Agentic Code Evolution
A recent study published on arXiv (2605.13874) presents GEAR (Genetic AutoResearch), a novel system that shifts from a single-path search method utilized by autonomous research agents to a population-based search across various research states. In contrast to traditional agents, which only refine one program while discarding valuable partial ideas and insights from unsuccessful experiments, GEAR preserves a collection of strong candidate solutions. It chooses parent solutions based on their productivity, novelty, and coverage, and fosters new ideas via mutation and crossover. Each research state archives code modifications, reflections, and performance metrics, allowing future decisions to leverage past findings. The paper examines three GEAR variants: one driven by prompts, one with a fixed search controller, and one with a learned controller, aiming to address the constraints of narrow searches in autonomous machine learning research.
Key facts
- GEAR stands for Genetic AutoResearch for Agentic Code Evolution.
- It replaces single-path search with population-based search over multiple research states.
- It keeps a set of strong candidate solutions.
- Parents are selected based on productivity, novelty, and coverage.
- New ideas are explored through mutation and crossover.
- Each research state stores code changes, reflections, and performance data.
- Three versions of GEAR are studied: prompting-controlled, fixed programmatic controller, and learned controller.
- The paper is available on arXiv with identifier 2605.13874.
Entities
Institutions
- arXiv