Research map / Large language models / Agents, code and math
Hierarchical Optimization via LLM-Guided Objective Evolution for Mobility-on-Demand Systems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a training-free hybrid where an LLM (DeepSeek-R1-Distill-Qwen-32B) evolves high-level MILP assignment objectives via harmony-search-driven prompt evolution, while Gurobi solves the low-level routing for mobility-on-demand dispatch.
Problem
Ride-hailing mobility-on-demand dispatch balancing spatially heterogeneous supply and demand; RL methods are data-hungry and hard to constrain, while decomposed optimization relies on manually designed high-level objectives that ignore low-level routing dynamics.
Approach
An LLM acts as a meta-optimizer that generates high-level assignment objectives at each simulation step from streamed vehicle/passenger states; a harmony search algorithm with three prompt-evolution operators evolves the prompts using simulator fitness, while a Gurobi-solved two-level MILP enforces constraints.
Key results
- Averaged about 16% waiting-time reduction over state-of-the-art baselines across scenarios (abstract claim).
- On Manhattan Scenario 9, the method achieves 4.01 min vs 6.82 min for Full RL and 8.98 min for EoH.
- On Chicago Scenario 9, the method achieves 15.37 min vs 19.79 min for EoH; closed-loop querying reduces waiting time to 2.24 min vs 4.62 min for single open-loop query in P130_C80_T1200.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the LLM-optimizer framework instead of FunSearch | Mobility-on-demand dispatch · New York taxi dataset, Scenario 9 (P200_C100_T1200) | mean passenger waiting time: 8.98 -> 4.01 | Helped |
| use the LLM-optimizer framework instead of EoH | Mobility-on-demand dispatch · Chicago taxi dataset, Scenario 9 (P200_C100_T1200) | mean passenger waiting time: 19.79 -> 15.37 | Helped |
| replace the Data block with the Model blueprint block (ablation) | LLM objective-generation prompt · P50_C30_T300 | mean passenger waiting time: 9.93 ± 0.00 -> 8.72 ± 0.91 | Helped |
| replace the Data block with the Model blueprint block (ablation) | LLM objective-generation prompt · P70_C60_T600 | mean passenger waiting time: 5.86 ± 0.00 -> 2.86 ± 0.29 | Helped |
| replace the Data block with the Model blueprint block (ablation) | LLM objective-generation prompt · P100_C80_T900 | mean passenger waiting time: 5.29 ± 0.00 -> 3.07 ± 0.89 | Helped |
| replace the Data block with the Model blueprint block (ablation) | LLM objective-generation prompt · P130_C80_T1200 | mean passenger waiting time: 6.46 ± 0.00 -> 4.54 ± 0.36 | Helped |
| add Gurobi-compatible constraints to the prompt (ablation) | LLM objective-generation prompt · P50_C30_T300 | mean passenger waiting time: 8.72 ± 0.91 -> 8.16 ± 1.48 | Helped |
| add Gurobi-compatible constraints to the prompt (ablation) | LLM objective-generation prompt · P70_C60_T600 | mean passenger waiting time: 2.86 ± 0.29 -> 4.12 ± 0.83 | Hurt |
| add Gurobi-compatible constraints to the prompt (ablation) | LLM objective-generation prompt · P100_C80_T900 | mean passenger waiting time: 3.07 ± 0.89 -> 2.59 ± 0.52 | Helped |
| add Gurobi-compatible constraints to the prompt (ablation) | LLM objective-generation prompt · P130_C80_T1200 | mean passenger waiting time: 4.54 ± 0.36 -> 2.25 ± 0.05 | Helped |
| query the LLM at each timestep using updated environment feedback instead of querying once (ablation) | LLM-optimizer framework · P130_C80_T1200 | mean passenger waiting time: 4.62 -> 2.24 | Helped |
| use high harmony memory consideration rate (HMCR = 0.9) (ablation) | Harmony search prompt evolution · P100_C80_T900 | optimization cost | Helped |
Surprising findings
- FunSearch beats the proposed method in the smallest Manhattan scenarios (e.g., 0.77 vs 1.55 min in Scenario 1), so the hybrid does not dominate everywhere.
- Without the restriction block, LLM-generated objectives fail at high rates (100% in some prompt settings), and invalid objectives fall back to a static default objective.
Limitations
- Relies on a 32B LLM and Gurobi solver; LLM outputs can be invalid (error rates up to 36.7% without constraint prompts).
- Evaluated only on synthetic scenarios derived from taxi statistics, using simplified simulator without congestion or stochastic travel times.
- Results averaged over only 3 runs with substantial variance in some settings.
Experimental setup
Datasets: New York TLC taxi trip records (Jan-Jun 2024) Chicago taxi trip records (Jan-Jun 2023)
Benchmarks: 9 synthetic NYC Manhattan downtown scenarios 9 synthetic Chicago central district scenarios
Metrics: mean passenger waiting time (min)
Baselines: Distance manual objective Distance x Utilization Temporal x Utilization Distance x Temporal x Utilization Default + RL-Seq Full RL FunSearch EoH
Builds on: FunSearch EoH harmony search
Closest papers by meaning
- One Step is Enough: Multi-Agent Reinforcement Learning Based on One-Step Policy Optimization for Order Dispatch on Ride-Sharing Platforms (NeurIPS 2026 · similarity 0.87)
- NondBREM: Nondeterministic Offline Reinforcement Learning for Large-Scale Order Dispatching (AAAI 2024 · similarity 0.86)
- GARLIC: GPT-Augmented Reinforcement Learning with Intelligent Control for Vehicle Dispatching (AAAI 2025 · similarity 0.85)
- Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- URB - Urban Routing Benchmark for RL-equipped Connected Autonomous Vehicles (NeurIPS 2025 · similarity 0.83)
- Divide and Conquer: Grounding LLMs as Efficient Decision-Making Agents via Offline Hierarchical Reinforcement Learning (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.