RefineEvo is a novel evolutionary framework for Automatic Heuristic Design that transforms the process from static trial-and-error into a planning-guided, experience-driven system. It utilizes a Planner to dynamically schedule evolutionary operators and a Reflector to distill lessons into a Bidirectional Experience Pool containing positive insights and negative pitfalls.

  • The Planner adapts search tools to evolving problem complexity based on the current state.
  • The Reflector populates a pool with trajectory-aware, situation-conditioned insights for reuse.
  • Experiments show consistent outperformance of strong baselines on classic combinatorial optimization benchmarks.
  • The system achieves superior solution quality while improving token efficiency.

RefineEvo enables more efficient and autonomous heuristic design by leveraging historical search experience to guide generation.