Bridging Policy and Real-World Dynamics: LLM-Augmented Rebalancing for Shared Micromobility Systems

This paper introduces AMPLIFY, an LLM-augmented framework that dynamically adapts shared micromobility rebalancing strategies in real-time to handle emergent events, thereby outperforming traditional methods in demand satisfaction and revenue on real-world data.

The Gist

Imagine a city filled with hundreds of electric scooters and bikes, ready to whisk people to work, school, or the park. This is a Shared Micromobility System. It sounds great, but it has a major headache: Rebalancing.

Think of these scooters like a deck of cards. If everyone in the city wants to go to the stadium at 5 PM, all the scooters end up at the stadium, and the downtown offices are empty. If it rains, everyone wants a scooter downtown, but they're all stuck at the stadium. The system needs to constantly shuffle the "cards" (scooters) to where they are needed.

The Problem: When the Unexpected Happens

For years, computer scientists have built smart algorithms to shuffle these scooters. These algorithms are like experienced chess players. They are great at predicting the "average" game. They know that on a Tuesday morning, people go from suburbs to the city center. They plan perfectly for that.

But what happens when the unexpected strikes?

• A massive music festival suddenly starts (a Demand Surge).
• A truck hits a cluster of scooters, taking 20 out of service (a Supply Crash).
• The city mayor suddenly says, "We need to make sure the poor neighborhoods get just as many scooters as the rich ones" (a Policy Change).

The old "Chess Player" algorithms get confused. They are too rigid. They either ignore these surprises or become so cautious that they stop working well even on normal days.

The Solution: AMPLIFY (The "Smart Co-Pilot")

The authors of this paper created a new system called AMPLIFY. Instead of replacing the old chess player, they gave it a Super-Intelligent Co-Pilot powered by a Large Language Model (LLM)—the same kind of AI that powers chatbots like me.

Here is how AMPLIFY works, using a simple analogy:

1. The Base Strategy (The Chess Player)

The system starts with a standard plan generated by a traditional computer program. "Move 5 scooters from Zone A to Zone B." This is the baseline.

2. The Co-Pilot (The LLM)

Suddenly, a text message comes in: "Hey, there's a huge crowd forming near the stadium! Also, the city wants us to be fairer to the south side."

The old computer program doesn't understand "huge crowd" or "fairness." It just sees numbers. But the LLM Co-Pilot is like a human manager who reads the news, understands the context, and speaks human language. It reads the message and says, "Oh, I see! The standard plan won't work here. We need to change it."

3. The "Self-Reflection" (The Double-Check)

This is the secret sauce. The LLM doesn't just guess. It acts like a careful editor.

• Draft: It suggests a new plan: "Move 20 scooters to the stadium."
• Self-Check: It pauses and asks itself: "Wait, do we even have 20 scooters there? If I move them, will the south side be empty? Did I break the city's fairness rule?"
• Revision: It realizes it made a mistake. It rewrites the plan: "Okay, let's move 15 to the stadium and 5 to the south side to keep things fair."

This "Self-Reflection" loop happens in seconds, ensuring the new plan is actually possible and follows the rules.

Why This is a Big Deal

The researchers tested this on real data from Chicago's e-scooters. Here is what they found:

• The "Normal" Days: The new system works just as well as the old, smart algorithms.
• The "Chaos" Days: When a festival happened or scooters broke down, the old systems crashed (satisfaction dropped to ~60%). The new AMPLIFY system stayed calm and effective (staying above 90%).
• The "Fairness" Test: When the city asked for more fairness, the old systems couldn't understand the request. The LLM understood the concept of "fairness" immediately and adjusted the scooters to match.

The Takeaway

Think of AMPLIFY not as a robot that replaces human thinking, but as a bridge. It connects the rigid, mathematical world of computer algorithms with the messy, unpredictable, "human" world of real-life events.

It takes the best of both worlds: the speed of a computer and the common sense of a human manager. By letting an AI "read the room" and adjust the plans on the fly, cities can keep their scooters where people actually need them, even when life throws a curveball.

Technical Summary

1. Problem Statement

Shared micromobility systems (e.g., e-scooters, bikes) face a critical operational challenge: vehicle rebalancing. Operators must redistribute vehicles across urban regions to match spatiotemporal demand.

• Limitations of Existing Methods:
  • Uncertainty-Ignorant Approaches: Rely on average demand patterns or pre-trained models (RL, optimization). They fail when "emergent events" occur (e.g., sudden demand surges, vehicle outages, or new regulatory constraints).
  • Uncertainty-Aware Approaches: Use robust optimization or learning-based strategies to handle predefined uncertainties. However, they often suffer from a robustness-performance trade-off, becoming overly conservative and degrading performance under normal conditions.
  • Human-in-the-Loop: While effective, manual intervention is not scalable due to resource constraints.
• The Gap: There is a lack of flexible, scalable solutions that can adapt pre-optimized policies in real-time to unforeseen emergent scenarios without requiring model retraining or encoding specific rules for every possible event.

2. Methodology: The AMPLIFY Framework

The authors propose AMPLIFY, a modular framework that augments existing rebalancing policies with a Large Language Model (LLM) agent to handle emergent scenarios.

Core Architecture

The system consists of two decoupled components:

1. Baseline Rebalancing Module: Generates an initial strategy ($a_t$) using standard methods (e.g., Reinforcement Learning, Genetic Algorithms, or rule-based optimization) based on current vehicle distribution and predicted demand.
2. LLM-Augmented Adaptation Module: Acts as a decision-making assistant that modifies the baseline strategy ($\hat{a}_t$) in response to emergent conditions ($E_t$).

Key Mechanisms

• Input Modality: The LLM accepts natural language descriptions of emergent situations (e.g., "a large crowd is forming near the stadium," "Region 3 vehicles are under maintenance," or "regulators require better equity"). This allows the system to ingest unstructured, real-world reports.
• Structured Prompting: The LLM is provided with:
  • System background and constraints.
  • Real-time status (vehicle distribution, predicted demand flows, historical variance).
  • The initial rebalancing strategy generated by the baseline module.
  • Specific emergent directives.
• Self-Reflection Loop: To ensure feasibility and correctness, the LLM employs an iterative self-reflection mechanism:
  1. Generate an initial adaptation.
  2. Critically evaluate the output against constraints (e.g., total vehicle count conservation, non-negative allocations, goal alignment).
  3. Revise and finalize the strategy.
  • Result: This loop significantly reduces invalid outputs (e.g., malformed formats or constraint violations).

3. Key Contributions

1. Novel Framework: Introduction of AMPLIFY, an LLM-augmented policy adaptation framework that enables real-time strategy adjustment with minimal computational overhead, avoiding the need for retraining.
2. Modular Design: Decoupling the "backbone" (the original optimization/RL algorithm) from the "adapter" (the LLM). This ensures compatibility with diverse existing rebalancing algorithms.
3. Grounding Strategy: Demonstrating that grounding the LLM with a pre-generated baseline strategy is superior to using the LLM for open-ended planning from scratch.
4. Comprehensive Evaluation: Extensive testing on real-world Chicago e-scooter data covering three emergent scenarios: Rising Demand, Shrinking Supply, and Dynamic Goals (Equity).

4. Experimental Results

The framework was evaluated using a dataset of 629,000+ e-scooter trips from Chicago (Lime, Spin, Bird).

• Performance under Rising Demand:
  • As demand surged by 100%, baseline methods (SDSM, GA, RECOMMEND) saw satisfaction rates drop significantly (e.g., SDSM dropped to ~61%).
  • AMPLIFY maintained a stable satisfaction rate of 96.87%, outperforming the best baseline by a significant margin.
• Performance under Shrinking Supply:
  • When fleet availability was reduced by 20%, AMPLIFY achieved 89.38% satisfaction, compared to 80.04% for the best baseline (RECOMMEND).
• Dynamic Goals (Equity):
  • When tasked with improving vehicle deployment equity, AMPLIFY improved the equity metric by 52.22% compared to the baseline, demonstrating the LLM's ability to interpret and optimize for complex, non-revenue goals.
• Ablation Studies:
  • Baseline Dependency: The LLM performs best when the underlying baseline strategy is strong (e.g., RL-based). If the baseline is weak (rule-based), the LLM cannot fully compensate for fundamental planning flaws.
  • LLM Choice: Larger models (GPT-4o, LLaMA-70B) significantly outperformed smaller models (GPT-3.5, LLaMA-8B), which often failed to generate valid strategies.
  • Self-Reflection: Enabling self-reflection increased the rate of valid, constraint-satisfying strategies from 15% to 85%.

5. Significance and Conclusion

This work bridges the gap between rigid, pre-optimized transportation policies and the fluid reality of urban dynamics.

• Flexibility: It allows operators to adapt to unforeseen events (like festivals or maintenance) using natural language inputs without retraining complex models.
• Scalability: By decoupling the adaptation logic from the core algorithm, it offers a generalizable solution for various micromobility systems.
• Practical Impact: The study proves that LLMs, when used as "adapters" rather than "planners from scratch," can effectively enhance system revenue, user satisfaction, and operational equity in real-world shared mobility networks.

In summary, AMPLIFY demonstrates that integrating LLMs with traditional optimization frameworks creates a robust, human-like decision-making layer capable of managing uncertainty in shared micromobility systems.