CogMCTS: A Novel Cognitive-Guided Monte Carlo Tree Search Framework for Iterative Heuristic Evolution with Large Language Models

2025-12-09

AI Summary

Existing LLM-driven Automated Heuristic Design (AHD) methods suffer from premature convergence to local optima, weak cognitive integration, and insufficient search diversity. To address these limitations, we propose CogMCTS—a novel framework integrating large language models’ cognitive mechanisms with Monte Carlo Tree Search (MCTS). Our approach introduces, for the first time, a multi-round cognitive feedback mechanism that synthesizes historical experience, node-state information, and negative outcomes, alongside a dual-track node expansion strategy balancing exploration and exploitation. Furthermore, we incorporate elite heuristic management and strategic mutation to enhance solution diversity. Experimental results demonstrate that CogMCTS significantly outperforms state-of-the-art AHD methods in solution quality, stability, and optimization efficiency across multiple combinatorial optimization benchmarks. By grounding heuristic search in interpretable, iterative cognitive principles, CogMCTS establishes a new paradigm for LLM-augmented combinatorial optimization.

Technology Category

• Search and Optimization: Heuristic Search
• Cognitive Modeling & Cognitive Systems: Agent Architectures
• Constraint Satisfaction and Optimization: Search

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Agentic search

Abstract

Automatic Heuristic Design (AHD) is an effective1 framework for solving complex optimization prob-2 lems. The development of large language mod-3 els (LLMs) enables the automated generation of4 heuristics. Existing LLM-based evolutionary meth-5 ods rely on population strategies and are prone6 to local optima. Integrating LLMs with Monte7 Carlo Tree Search (MCTS) improves the trade-off8 between exploration and exploitation, but multi-9 round cognitive integration remains limited and10 search diversity is constrained. To overcome these11 limitations, this paper proposes a novel cognitive-12 guided MCTS framework (CogMCTS). CogMCTS13 tightly integrates the cognitive guidance mecha-14 nism of LLMs with MCTS to achieve efficient au-15 tomated heuristic optimization. The framework16 employs multi-round cognitive feedback to incor-17 porate historical experience, node information, and18 negative outcomes, dynamically improving heuris-19 tic generation. Dual-track node expansion com-20 bined with elite heuristic management balances the21 exploration of diverse heuristics and the exploita-22 tion of high-quality experience. In addition, strate-23 gic mutation modifies the heuristic forms and pa-24 rameters to further enhance the diversity of the so-25 lution and the overall optimization performance.26 The experimental results indicate that CogMCTS27 outperforms existing LLM-based AHD methods in28 stability, efficiency, and solution quality.

Problem

• Enhances automated heuristic design for complex optimization problems
• Integrates LLM cognitive guidance with MCTS to improve exploration-exploitation balance
• Addresses local optima and limited search diversity in LLM-based evolution

Innovation

• Integrates LLM cognitive guidance with Monte Carlo Tree Search
• Uses multi-round feedback and dual-track node expansion
• Applies strategic mutation to enhance solution diversity