LLM-Assisted MCTS Framework for Large-Scale CVRP Solver Design
A new framework called LaF-MCTS (LLM-assisted Flexible Monte Carlo Tree Search) automates the design of solvers for large-scale Capacitated Vehicle Routing Problems (LSCVRP) with hundreds to thousands of nodes. Traditional divide-and-conquer approaches require expert-crafted decomposition logic and sub-solver configuration, which is labor-intensive. LaF-MCTS uses a three-tier decision hierarchy to incrementally design decomposition policies and sub-solvers, overcoming the limited context window of LLMs. The framework is detailed in arXiv preprint 2605.03339.
Key facts
- LaF-MCTS automates LSCVRP solver design
- Uses three-tier decision hierarchy
- Addresses LLM context window limitations
- Published on arXiv as 2605.03339
- Targets problems with hundreds to thousands of nodes
- Combines LLM with Monte Carlo Tree Search
- Eliminates need for expert decomposition logic
- Framework is novel and automated
Entities
Institutions
- arXiv