GRAFT: Execution-Guided Genetic Recombination for Agent Design
Summary
- Researchers introduced Execution Guided Genetic Recombination for Agent Design (GRAFT), an evolutionary method for LLM-based agent design.
- Unlike standard approaches relying solely on aggregate fitness, GRAFT utilizes parent-specific execution evidence to guide semantic crossover and integrated mutation.
- The feedback harness reuses bounded samples of successful and failed runs without requiring additional task executions.
- Evaluated across six reasoning and programming settings, GRAFT achieved a mean test accuracy of 82.0%, outperforming the strongest baseline by 9.8 percentage points.
Why it matters
GRAFT improves automated LLM agent design by leveraging instance-level execution evidence rather than aggregate scores alone, significantly boosting performance in reasoning and programming tasks.