Last-Mile Routing: Where ML Outperforms LLMs (and Where LLMs Pull Ahead)

Challenges in Last-Mile Routing Explained Last-mile delivery represents the final frontier in logistics, accounting for up to 50% of total supply chain costs due to its complexity. Vehicle Routing Problems (VRPs) here involve dynamic constraints like time windows, vehicle capacities, traffic variability, customer priorities, and real-time disruptions such as weather or demand surges. Traditional optimization solvers struggle with NP-hard scalability, leading to heuristics that trade off optimality for speed. Enter AI: Machine Learning (ML) has long powered route optimization machine learning tools, while Large Language Models (LLMs) promise natural language interfaces for AI last-mile delivery. Yet, supply chain routing AI must balance feasibility, efficiency, and adaptability without regressions in service levels. Key challenges include: - Constraint satisfaction : Ensuring routes respe

ct hard limits like delivery windows and load limits. - Scalability : Optimizing for 100s–1000s of stops daily across fleets. - Uncertainty handling : Incorporating probabilistic elements like traffic or no-shows. - Interpretability : Explaining decisions to dispatchers and customers. For B2B leaders, the question is: Does LLM hype justify replacing proven ML for last mile optimization? Where Traditional ML Excels Over LLMs Traditional ML, especially reinforcement learning (RL) and graph neural networks (GNNs), remains the gold standard for core last-mile tasks. These models are trained on vast logistics datasets to solve ML VRPs logistics problems with near-optimal solutions at scale. Proven Scalability and Optimality ML frameworks like OR-Tools with ML enhancements or specialized solvers (e.g., Google OR-Tools + RL) handle constraint-heavy VRPs efficiently. For instance, RL agents in V

RPAGENT achieve 5-10% better optimality gaps than classical heuristics on benchmarks like CVRPTW (Capacitated VRP with Time Windows). LLMs falter here due to hallucinations in numerical reasoning and poor handling of combinatorial explosion. A study on arXiv ( ) shows LLMs underperform ML in time interval predictions critical for routing, even with extended context. Real-World Failure Modes of LLMs - Feasibility violations : LLMs often propose invalid routes ignoring capacities (e.g., overloading vans). - Scalability limits : Prompting GPT-4o or Claude 3.5 Sonnet for 500-stop VRPs leads to timeouts or degraded quality. - Quantitative weaknesses : Struggles with temporal structures, per the same arXiv paper. In routine operations, smaller ML models (SLMs) are faster and cheaper, as noted in . LLM Strengths in Adaptive Routing Scenarios LLMs shine where flexibility trumps raw optimization.

Their natural language processing enables LLMs vehicle routing for unstructured inputs, like parsing customer notes or integrating multi-modal data (e.g., weather APIs + emails). Adaptive and Strategic Use Cases - Exception handling : Repurposing delayed packages via conversational rerouting ("Reschedule this high-priority delivery around traffic?"). - Heuristic generation : LLMs auto-generate custom rules, e.g., prioritizing VIPs based on sentiment analysis of orders. - Cross-domain generalization : An arXiv paper ( ) highlights LLMs improving forecasting in novel scenarios, transferable to dynamic routing. For strategic planning, LLMs interpret ML outputs, bridging automation and human oversight ( ). Key Benchmarks and Research Comparisons Benchmarks underscore ML's edge: Benchmark ML Performance LLM Performance Source ----------- ---------------- ----------------- -------- CVRPTW (10

0 stops) 2-5% optimality gap 15-30% gap VRPAGENT papers Dynamic VRP RL: 95% feasibility LLMs: 70-80% arXiv:2601.10132 Time Series Routing SLMs/ML: High accuracy LLMs: Context-dependent Analytics Insight These are aggregated from public arXiv/SERP sources as of 2026-05 (e.g., no invented pricing; focused on relative gaps). ML dominates scalable VRPs; LLMs lag in feasibility but excel in zero-shot adaptation. Hybrid Approaches: Combining ML and LLMs The future lies in hybrids: ML for optimization, LLMs for orchestration. - ML Core + LLM Wrapper : Use ML solvers (e.g., GNN-RL) for routes, LLMs for natural language queries and refinements. - Multi-Agent Systems : LLMs as coordinators in frameworks like VRPAGENT, delegating sub-VRPs to ML agents. - Prompt Engineering : Feed ML outputs to LLMs for explanation or tweaks, reducing hallucinations via chain-of-thought. Research shows hybrids boost

performance 10-20% in mixed scenarios ( ). Real-World Implementation with LUMOS Platform LUMOS, a multi-agent platform, exemplifies hybrid power for last-mile. It integrates ML VRP solvers with LLM agents for adaptive logistics: - VRPAGENT Integration : ML agents solve core routing; LLM agents hand