MeEvo: Metacognitive Evolution Combined with Natural Evolution for Automatic Heuristic Design

Researchers have proposed MeEvo, a dual-layer framework for automatic heuristic design that couples natural evolution with metacognitive evolution to improve the stability and quality of machine-generated heuristics, according to a paper posted to arXiv on 12 June 2026 [1]. Large language models have enabled automatic heuristic design by synthesizing reasoning traces and code, but existing architectures fall into two camps, each with limitations [2]. Natural evolution uses crossover and mutation to explore heuristic programs but discards reasoning traces, weakening knowledge inheritance. Metacognitive evolution refines reasoning through reflection but lacks population-level recombination, which can lead to premature convergence [2]. MeEvo addresses these gaps by cycling between the two modes. Natural evolution explores heuristic code while recording 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 for the next cycle [2]. The framework was tested on five optimization problems using two LLM backbones and achieved stronger and more stable performance than existing LLM-based AHD architectures, particularly on complex constrained tasks [2]. The authors note that the cyclical design allows population-driven exploration and reflection-driven refinement to reinforce each other [2]. The paper appears on arXiv under the Computer Science category for Neural and Evolutionary Computing and was submitted on 12 June 2026 [1]. The work builds on a broader trend of using LLMs to automate heuristic generation, a subfield that has drawn attention for its potential to reduce manual engineering effort in optimization [2].