Large Language Model-Driven Full-Component Evolution of Adaptive Large Neighborhood Search

2026-03-10 · Source: cs.NE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, extended

Summary

A new closed-loop, large language model (LLM)-driven evolutionary framework has been developed to automate the design of Adaptive Large Neighborhood Search (ALNS) algorithms, a prominent metaheuristic for production and logistics optimization. This framework decouples ALNS into seven core modules: destroy, repair, operator selection, weight update, initial solution construction, acceptance rule, and destroy-rate control, evolving each independently. By incorporating the MAP-Elites mechanism, it maintains a multi-dimensional elite archive to drive the evolution of both solution quality and strategic diversity. Tested on TSPLIB benchmarks, the evolved algorithms consistently outperform optimized classic ALNS baselines under fixed-iteration and fixed-time limits. Notably, on large-scale instances, the average optimality gap drops from 3.18% to 0.74%. The study also reveals counterintuitive yet meaningful design patterns that emerged during evolution and compares the effectiveness of different LLMs (GPT-5.2, Grok-Code, MiniMax-m2, DeepSeek-v3.2) in supporting this automated algorithm design.

Key takeaway

For research scientists developing optimization algorithms, this work demonstrates that LLM-driven automated design can significantly outperform human-expert baselines. You should consider adopting a decoupled, evolutionary framework for complex metaheuristics like ALNS, focusing on evolving core components such as destroy/repair operators and destruction-degree control. This approach not only yields superior solution quality and computational efficiency but also uncovers novel, non-intuitive design patterns that can inform future algorithm development.

Key insights

LLM-driven evolution can automate and enhance complex metaheuristic algorithm design by decoupling and optimizing core components.

Principles

• Decouple complex algorithms into evolvable modules.
• Use quality-diversity archives to maintain diverse solutions.
• Dynamic context injection strengthens LLM learning.

Method

The framework uses a "generate–evaluate–feedback" loop where an LLM acts as an intelligent mutator. It employs a fixed instruction template with dynamic context injection and an isolated evaluator for each component, guided by a multi-objective quality model.

In practice

• Evolved destroy/repair operators significantly improve ALNS performance.
• Dynamic destruction-degree control is critical for balancing search.
• Momentum-based Bayesian Bandit can enhance operator selection.

Topics

• Adaptive Large Neighborhood Search
• Large Language Models
• Automated Algorithm Design
• Program Evolution
• Traveling Salesman Problem

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• LLM-Driven Co-Evolutionary Automated Heuristic Design for Bi-Component Coupled Combinatorial Optimization — CoEvo-AHD uses LLMs and co-evolution to design heuristics for coupled combinatorial optimization problems, improving solution quality.
• What Makes an LLM a Good Optimizer? A Trajectory Analysis of LLM-Guided Evolutionary Search — Strong LLM optimizers act as local refiners, making incremental improvements within localized semantic search spaces.
• Solvita: Enhancing Large Language Models for Competitive Programming via Agentic Evolution — Solvita enhances LLMs for competitive programming through a multi-agent, reinforcement learning framework with continuous, experience-driven knowledge networks.
• Agentic Environment Engineering for Large Language Models: A Survey of Environment Modeling, Synthesis, Evaluation, and Application — Agentic environment engineering for LLMs involves systematic modeling, synthesis, evaluation, and co-evolution to enhance model capabilities.
• REFLEX: Reflective Evolution from LLM Experience — REFLEX decouples visual diagnosis from code generation in LLM-guided evolutionary search for transparent, efficient policy discovery.
• Solving the Wrong Problem Works Better - Robert Lange — Shinka Evolve uses LLMs and evolutionary algorithms for sample-efficient program generation and scientific discovery.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.NE updates on arXiv.org.