Large Language Models as Delivery Rider: Generating Instant Food Delivery Riders' Routing Decision with LLM Agent Framework

1. Problem Statement

The paper addresses the challenge of modeling the complex, heterogeneous decision-making behaviors of delivery riders in on-demand instant delivery systems.

• Limitations of Current Models: Traditional agent-based models (ABM) and operations research approaches often rely on simplified, homogeneous rules (e.g., "accept the nearest order" or "maximize immediate profit"). These fail to capture the nuanced cognitive processes, strategic reasoning, and diverse work patterns (personas) of real-world riders.
• Data Scarcity: Fine-grained rider routing data is often proprietary or scarce, making empirical analysis of strategic interactions difficult.
• The Gap: There is a lack of frameworks capable of simulating cohort-level interactions in dynamic, competitive environments where agents must make strategic routing decisions based on specific behavioral priors (e.g., risk aversion, income maximization, peak-hour specialization).

2. Methodology: The LLM-DR Framework

The authors propose LLM-DR, a novel agent framework that combines data-driven persona extraction with Large Language Model (LLM) reasoning to simulate rider behavior. The framework operates through four key stages:

A. Empirically-Grounded Persona Extraction

Instead of manually defining agent rules, the authors derive personas from real-world data:

• Data Source: A large-scale, real-world trajectory dataset of delivery riders.
• Technique: Unsupervised clustering (k-means) on 15 behavioral metrics (e.g., working hours, order stacking frequency, income volatility, travel distance).
• Result: Four distinct rider personas were identified:
  1. Full-time Workhorse: Consistent activity throughout the day; accepts a wide variety of orders for steady income.
  2. Lunch-Peak Specialist: Highly active during 11:00–14:00; focuses on short-distance bundles for peak efficiency.
  3. Super Stacker: Highly selective; rejects single orders to wait for complex, multi-order "waves" to maximize earnings per trip.
  4. Dinner-Peak Core: Prefers simple, single-order trips during the evening rush (17:00–21:00).

B. High-Fidelity Simulation Environment

• Discrete-Time Simulation: The system runs in 10-minute intervals.
• Order Broadcasting: New orders are broadcast to all idle agents within a 2km radius of the pickup location, mimicking real platform "order pooling."
• Dynamic Rewards: Fees are calculated based on distance (e.g., ¥3 for <1km, up to ¥15 for >5km).

C. Chain-of-Thought (CoT) Decision-Making

Each persona is instantiated as an LLM agent (using Gemini-2.5-Flash) that employs a structured reasoning process:

1. Perception: The agent analyzes available orders (location, fee, time).
2. Evaluation: The agent evaluates orders against its specific persona constraints (e.g., a "Super Stacker" looks for shared pickup locations).
3. Planning: The agent formulates a route and sequence for multiple orders.
4. Decision: The agent outputs a final decision (accept/reject) and a justification.

D. Simulation Loop

The environment updates the agent's status (IDLE/DELIVERING), calculates routes via mapping APIs (Amap), updates locations, and logs earnings. The cycle repeats for a 24-hour simulation period.

3. Key Contributions

1. Novel Framework (LLM-DR): The first framework to integrate unsupervised clustering of real-world trajectory data with LLM agents to simulate heterogeneous rider strategies in a dynamic market.
2. Data-Driven Personas: Moves beyond theoretical agent definitions by grounding agent behaviors in four empirically derived clusters, ensuring behavioral realism.
3. Reasoning-Based Routing: Introduces a Chain-of-Thought mechanism that allows agents to simulate the cognitive deliberation behind routing choices, not just the action itself.
4. Interactive Market Simulation: Successfully models high-frequency, strategic interactions where agents compete for orders in real-time, a significant step up from static or pre-planned travel diary generation.

4. Experimental Results

The framework was validated using a real-world order dataset from Beijing (9,025 orders over 24 hours). Four agents (one per persona) were simulated from different starting locations (Zhongguancun, Xidan, Wangjing, Sanlitun).

• Behavioral Consistency: The LLM agents exhibited distinct behaviors matching their personas:
  • Lunch-Peak Specialist rejected orders outside peak hours.
  • Super Stacker frequently rejected single orders to wait for multi-order opportunities.
  • Full-time Workhorse maintained consistent activity.
• Mobility Patterns:
  • Spatial: The Workhorse covered the broadest area; the Lunch Specialist was concentrated in business districts; the Dinner Core focused on residential areas.
  • Temporal: Clear clustering of activity around specific peak times (noon vs. evening) or continuous operation.
• Performance Metrics (Averaged):
  • Full-time Workhorse: Highest total earnings (¥118.75) and mileage (110.15 km), accepting the most orders (24).
  • Lunch-Peak Specialist: High efficiency (lowest mileage per order: 3.15 km) with earnings of ¥103.00.
  • Super Stacker: Moderate earnings (¥101.75) with high strategic complexity (accepting fewer, but more complex, orders).
  • Dinner-Peak Core: Lowest earnings (¥94.67) but focused on evening simplicity.

5. Significance and Future Work

• Scientific Impact: This work bridges Computational Social Science and Operations Research, demonstrating that LLM agents can replicate complex human strategic reasoning in competitive environments. It shifts the paradigm from simulating "what people do" to "why they do it."
• Practical Application: The framework serves as a tool for urban planners and platform operators to:
  • Analyze system-level outcomes (traffic congestion, labor distribution).
  • Test "what-if" scenarios (e.g., impact of extreme weather, sudden demand surges, or policy changes).
• Future Directions: The authors plan to simulate extreme weather events and demand shocks to analyze how different personas adapt to stress and volatility, further refining the understanding of urban logistics resilience.

In summary, LLM-DR provides a pioneering methodology for generating high-fidelity simulations of human mobility in on-demand delivery, proving that LLMs can effectively model the strategic heterogeneity of human workers in dynamic systems.