MeEvo combines natural and metacognitive evolution for better heuristic design
LLMs design heuristics faster and more reliably with a dual-layer evolution framework
MeEvo addresses limitations in existing LLM-based automatic heuristic design (AHD) paradigms. Natural evolution approaches discard reasoning traces, weakening knowledge inheritance, while metacognitive evolution lacks population-level recombination, risking premature convergence. MeEvo's dual-layer design solves both: natural evolution explores heuristic code and records reasoning traces, fitness values, and errors into a shared history; metacognitive evolution then reflects on that history to generate improved heuristics that re-enter the parent pool. This cyclical coupling enables population-driven exploration and reflection-driven refinement to reinforce each other.
Experiments across five optimization problems using two different LLM backbones demonstrate that MeEvo consistently outperforms existing LLM-based AHD frameworks in both performance and stability, particularly on complex constrained optimization tasks. The framework's ability to combine the strengths of both evolution paradigms marks a significant step forward in automated heuristic design, potentially accelerating solutions for logistics, scheduling, and engineering optimization problems.
- MeEvo uses a dual-layer framework combining natural evolution (crossover/mutation) and metacognitive evolution (reflection on shared history).
- Evaluated on five optimization problems with two different LLM backbones, showing stronger and more stable performance.
- Overcomes limitations of prior LLM-based AHD: reasoning traces are preserved, and population-level recombination prevents premature convergence.
Why It Matters
MeEvo's hybrid evolution approach could dramatically improve automated optimization in logistics, scheduling, and engineering.