ARTFEED — Contemporary Art Intelligence

Two-Stage AST-Based Evolutionary Operator Expands LLM-Driven Heuristic Design

ai-technology · 2026-04-22

A novel two-phase evolutionary operator for Large Language Model-based automated heuristic design (LLM-AHD) has been introduced to improve the exploration of algorithm search spaces. In the initial phase, Abstract Syntax Trees (ASTs) of heuristic code are altered, resulting in a variety of structural variants, many of which are invalid, thus challenging the validity constraints of traditional one-phase approaches. The subsequent phase involves the LLM converting these invalid heuristics into functional code. Existing LLM-AHD models restrict exploration by depending on valid code and the pre-existing knowledge of the LLM. This new approach distinguishes between structural variation and code repair, thereby broadening the search space. This study, referenced as arXiv:2604.16420v1, seeks to enhance the efficiency of heuristics identified through LLM-based methods.

Key facts

  • A two-stage evolutionary operator for LLM-based automated heuristic design has been proposed
  • The first stage performs crossover and mutation directly on Abstract Syntax Trees (ASTs) of heuristic code
  • This stage intentionally generates diverse but often invalid structural variants
  • The second stage employs the LLM to repair invalid heuristics into executable, high-quality code
  • The approach breaks validity-induced boundaries that limit algorithm search space exploration
  • Existing one-stage methods rely entirely on LLM's pre-trained knowledge and require valid code during operation
  • Conventional frameworks use semantic evolutionary operators and "thought-code" representation
  • The paper is identified as arXiv:2604.16420v1 and announced as cross-disciplinary research

Entities

Institutions

  • arXiv

Sources