ARTFEED — Contemporary Art Intelligence

TurboEvolve Framework Enhances LLM-Driven Program Evolution Efficiency

ai-technology · 2026-04-22

TurboEvolve, a novel multi-island evolutionary framework, tackles the challenges of cost and reliability in LLM-driven program evolution. By employing verbalized sampling, the system enhances sample efficiency and robustness within fixed evaluation budgets, prompting the LLM to produce a variety of candidates with self-assigned weights. An adaptive online scheduler modifies the candidate count to foster exploration during periods of stagnation while minimizing overhead during consistent progress. Seed-pool injection organizes existing solutions and disperses them across islands, utilizing controlled perturbations and elitist preservation to maintain a balance between diversity and refinement. TurboEvolve demonstrates superior performance at reduced budgets across various program-optimization benchmarks. This research was published on arXiv under identifier 2604.18607v1.

Key facts

  • TurboEvolve is a multi-island evolutionary framework for LLM-driven program evolution
  • It addresses cost and run-to-run variance issues in discovering high-quality programs
  • The framework introduces verbalized sampling to prompt LLMs to emit K diverse candidates with self-assigned weights
  • An online scheduler adapts K to expand exploration under stagnation and reduce overhead during steady progress
  • Seed-pool injection clusters seeds and assigns them across islands with controlled perturbations and elitist preservation
  • The approach balances diversity and refinement in program optimization
  • TurboEvolve achieves stronger performance at lower budgets across multiple benchmarks
  • The research was announced on arXiv with identifier 2604.18607v1

Entities

Institutions

  • arXiv

Sources