RouteRepair: Instance-Level Failure Diagnosis and Targeted Repair in LLM-Based Automated Heuristic Design for Routing Optimization

Routing optimization is a critical component of logistics and transportation, requiring high-quality heuristics to solve complex problems under strict time and computational constraints. While Large Language Models (LLMs) have recently been used to automate the design of these heuristics, they often rely on aggregate performance metrics. This approach can hide specific, recurring failures that occur only on certain types of routing instances. RouteRepair is a new framework designed to solve this by identifying these specific weaknesses and applying precise, targeted repairs to the heuristic without degrading the performance on instances that are already being solved effectively.

Diagnosing and Repairing Weaknesses

RouteRepair shifts the focus from general performance to instance-level diagnosis. By analyzing routing evidence, solver behavior, and the underlying program context, the system identifies exactly where a heuristic is failing. Once a weakness is pinpointed, RouteRepair defines a "bounded repair objective." This ensures that the modifications made by the LLM are narrow and specific, addressing the identified failure while intentionally protecting the parts of the code that are already functioning well. The ai search story also surfaces in Stanford AI discovery identifies natural weight..., adding another angle.

Validating Improvements

A key feature of RouteRepair is its rigorous validation process. Every time a repair is proposed, the system performs a matched parent-child evaluation. This compares the original heuristic (the parent) with the modified version (the child) across the same set of instances. This dual-check approach ensures that the repair actually fixes the intended failure while simultaneously monitoring for "collateral degradation"—the risk that fixing one problem might inadvertently break another part of the heuristic.

Performance Gains

The researchers tested RouteRepair across several routing challenges, including the Traveling Salesman Problem (TSP) and the Capacitated Vehicle Routing Problem (CVRP). The results demonstrate significant improvements:

• TSP Optimization: Using Guided Local Search (GLS), RouteRepair reduced the mean optimality gap from 1.7476% to 0.7587%.

• CVRP Efficiency: The constructive heuristic developed by RouteRepair lowered the average route cost by 1.91% compared to the standard savings heuristic.

• ACO Priors: The framework successfully generated Ant Colony Optimization (ACO) priors that outperformed hand-designed versions. The ai search story also surfaces in Google AI Releases TimesFM 3 for..., adding another angle.

A Failure-Aware Approach

The success of RouteRepair highlights the importance of "failure-aware" refinement. By moving away from broad, aggregate-based training and toward evidence-constrained, targeted modifications, the framework proves that it is possible to improve heuristics on difficult, specific problem structures without sacrificing the reliability of the system on cases it has already mastered. The ai search story also surfaces in OpenAI Says AI Found Possible Navier–Stokes..., adding another angle. as detailed in the full paper on Arxiv