💻 computer science

MARS-Agent: Retrieval-Augmented Multi-Agent Deliberation for Cross-Border E-Commerce Supply Chain Resilience

The paper proposes MARS-Agent, a retrieval-augmented multi-agent framework featuring a Dual-Triggered RAG mechanism and a Verify-Adapt-Replan Loop to enable closed-loop, self-correcting deliberation for enhancing cross-border e-commerce supply chain resilience against heterogeneous disruptions.

Original authors: Wenhui Zheng, Qing Xu

Published 2026-09-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wenhui Zheng, Qing Xu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Global trade has always been a delicate balancing act, but the modern version that moves goods across borders for online shopping faces a unique kind of pressure. Unlike a local delivery network that operates within a single country's stable laws and roads, a cross-border supply chain must navigate a shifting landscape of different governments, changing trade rules, and unpredictable political events. When a port gets clogged or a new sanction is announced, the ripple effects can be immediate and severe. For years, computer systems designed to manage these networks have operated like a one-way street: they would spot a problem, calculate a solution, and send it out. However, these systems often failed to realize that their own solutions could create new problems. If a manager reroutes a shipment to avoid a storm, that new route might suddenly run into a different country's customs blockade or a supplier who just got hit by a trade restriction. The old methods treated the initial risk and the consequences of the fix as separate events, leaving the system vulnerable to a chain reaction of errors that it could not see coming.

To solve this, researchers Wenhui Zheng and Qing Xu have developed a new system called MARS-Agent, which acts more like a team of specialists constantly checking their work against the real world. Instead of a single computer program making a decision and moving on, this system uses a group of artificial intelligence agents that collaborate in a continuous loop. The process begins with a "scout" that constantly scans the internet for news about geopolitical conflicts, weather, and trade policies, turning those messy news stories into clear facts. An "analyst" then translates those facts into numbers, estimating how much a delay might cost or how likely a shipment is to fail. An "optimizer" uses this information to draft a plan, such as choosing a new shipping route or a different warehouse. But the crucial difference in this new system is the final step: a "validator" agent that acts as a safety inspector. Before any plan is executed, this validator checks the proposed route and the new suppliers involved to see if they introduce fresh risks that were not present in the original problem. If the validator finds a hidden danger, it sends the plan back to the optimizer to be rewritten, creating a cycle of checking and refining until the plan is safe to use.

The researchers tested this approach in a highly detailed simulation of a cross-border e-commerce network, injecting it with 120 different real-world disruption scenarios ranging from port congestion caused by typhoons to sudden trade sanctions and geopolitical conflicts. They compared their new system against five other methods, including traditional mathematical models and other artificial intelligence techniques. The results showed that the MARS-Agent team consistently outperformed the others. In the simulations, it managed to deliver orders on time nearly 90 percent of the time, a significant improvement over the next best method. More importantly, it reduced the total cost of operations and the penalties for late deliveries by a wide margin. Perhaps most critically, the system lowered the overall risk exposure by more than a third compared to the best existing risk-aware methods. This success was not just about being faster or cheaper; it was about the system's ability to catch its own mistakes before they happened. In one specific test case involving a typhoon hitting a major port, older systems rerouted goods to a backup supplier only to have those goods rejected by customs because they missed a new export restriction. The MARS-Agent, however, caught this new restriction during its validation phase, forced a replanning, and found a different supplier that could actually fulfill the order without delay.

The study also looked at how much extra computing power this careful checking required. Because the system runs through multiple rounds of checking and re-checking, it takes longer to make a single decision than a standard program that acts immediately. The simulation showed that a full decision cycle took about 22 seconds, compared to just over 5 seconds for a simpler system. However, the researchers noted that in the real world of international shipping, decisions are often made over the course of hours or days, not seconds. This means the extra time spent verifying the plan is a small price to pay for avoiding the massive costs of a failed shipment. The system proved particularly effective when facing complex, unpredictable events like political conflicts, where the rules change rapidly and are hard to predict with standard formulas. By constantly pulling in fresh information and checking its own plans against that new data, the system avoided the trap of relying on outdated assumptions.

Ultimately, the work demonstrates that resilience in a chaotic global environment requires more than just reacting to problems; it requires a system that questions its own solutions. The MARS-Agent framework shows that by breaking down the decision-making process into specialized roles and forcing the system to verify its own plans against the latest information, it is possible to navigate the complexities of cross-border trade with far greater stability. The researchers found that this closed-loop approach, where the system constantly learns from its own proposed actions, is the key to keeping supply chains moving when the world around them is shifting. While the study was conducted in a simulated environment, the results suggest that this method of collaborative, self-correcting artificial intelligence could provide a robust foundation for managing the fragile and interconnected networks that keep global commerce flowing.

Technical Summary: MARS-Agent for Cross-Border E-Commerce Supply Chain Resilience

1. Problem Statement

Cross-border e-commerce supply chains face increasingly heterogeneous disruptions, including geopolitical uncertainties, trade policy shifts, and logistics bottlenecks. Existing methods for supply chain resilience suffer from a fundamental structural deficiency: they treat risk perception and operational decision-making as decoupled, open-loop processes.

This decoupling creates two critical issues:

Reactive Limitations: Classical methods (stochastic programming, robust optimization) rely on predefined disruption distributions and struggle with out-of-distribution events. Reinforcement learning (RL) approaches depend on structured state representations and fail to capture unstructured signals like geopolitical events.
Reflexive Blind Spots: Cross-border disruptions are inherently reflexive; a corrective action taken to mitigate one disruption may inadvertently introduce new risk exposures across interconnected jurisdictions. Current Large Language Model (LLM) and Multi-Agent systems operate in single-turn or open-loop settings, lacking mechanisms to verify whether their own decisions amplify risk.

2. Methodology: MARS-Agent

The authors propose MARS-Agent, a retrieval-augmented multi-agent collaboration framework designed to unify real-time risk perception, structured impact quantification, and closed-loop strategy optimization. The framework consists of three core components:

A. Dual-Triggered Retrieval-Augmented Generation (DT-RAG)

Unlike conventional RAG systems that rely on passive, query-driven retrieval, DT-RAG initiates retrieval proactively through two pathways:

State-Triggered Retrieval: Activated when supply chain state indicators (e.g., inventory levels, delay rates) breach specific thresholds. This allows the system to detect emerging external risks before they escalate.
Decision-Triggered Retrieval: Activated by the entities introduced in a candidate plan (e.g., new routes or backup suppliers). This addresses the "blind spot" where the risk implications of the system's own actions remain unexamined.
Data Ingestion: The system processes unstructured risk signals (from the GDELT Project) via a vector database and structured operational data (from the Olist dataset) via a relational database.

B. Four-Role Multi-Agent Architecture

The framework employs four specialized agents that collaborate to transform unstructured signals into executable strategies:

Environmental Scout Agent: Acts as a shared knowledge gateway, interfacing with the vector database to retrieve relevant risk records based on state or decision triggers.
Impact Analyst Agent: Uses LLM reasoning to convert retrieved unstructured text into structured disruption parameters (expected delay, extra cost, failure probability) with calibrated confidence intervals.
Optimization Agent: Generates resilience-aware supply chain plans (rerouting, supplier selection) by solving an objective function that balances cost and risk. It supports iterative replanning when new constraints are injected.
Validator Agent: Serves as the critical closed-loop component. It extracts new entities from a candidate plan, triggers decision-based retrieval, and quantifies the net risk change ( $\Delta Risk$ ).

C. Verify-Adapt-Replan Loop (VARL)

The VARL transforms the decision pipeline from feed-forward to a self-correcting deliberation process:

Forward Pass: The Scout, Analyst, and Optimization agents generate an initial plan ( $\pi_0$ ).
Verification: The Validator Agent evaluates $\pi_0$ against newly discovered risks. If the net risk change exceeds a threshold ( $\epsilon$ ), the plan is rejected.
Adaptation & Replanning: If rejected, the discovered risks are injected as constraints into the optimization objective, prompting the Optimization Agent to generate a revised plan ( $\pi_{n+1}$ ).
Convergence: The loop iterates until risk convergence, plan stability, or a maximum iteration bound ( $N_{max}$ ) is reached. If convergence fails, the system escalates to human review.

3. Key Contributions

The paper outlines four primary contributions:

MARS-Agent Framework: A unified four-role architecture (Scout, Analyst, Optimization, Validator) that integrates risk perception, impact quantification, and closed-loop optimization.
DT-RAG Mechanism: A proactive retrieval paradigm where knowledge acquisition is triggered by both supply chain state anomalies and agent-generated decisions, ensuring reasoning is anchored in up-to-date external knowledge.
VARL Mechanism: A self-correcting deliberation loop that stress-tests candidate plans through secondary retrieval and feeds discovered risks back as constraints, formally guaranteeing convergence.
Empirical Validation: Extensive experiments demonstrating superior performance across cost efficiency, delivery timeliness, and risk exposure compared to traditional optimization, RL, and LLM-based baselines.

4. Experimental Results

Experiments were conducted on a cross-border e-commerce simulation environment using GDELT geopolitical data and Olist order data, with 120 injected disruption scenarios (port congestion, trade sanctions, geopolitical conflicts, natural disasters).

Performance Metrics:

Cost Efficiency: MARS-Agent achieved a total cost of 108.2 and penalty cost of 18.6, outperforming the best baseline (PPO-Inv) by 13.3% and 31.4% respectively.
Delivery Timeliness: Achieved an 89.7% On-Time Delivery (OTD) rate with an average delay of 2.4 days, an 8.5% improvement in OTD over the best baseline.
Risk Exposure: Attained a risk score of 0.281, a 33.6% reduction compared to the best risk-aware baseline (SC-GPT).

Ablation Studies:

Removing the multi-agent architecture reduced OTD by 3.1% and increased risk.
Removing DT-RAG (reverting to passive retrieval) increased total cost by 7.8 and risk score by 0.071.
Removing the VARL (open-loop execution) resulted in the most significant degradation, increasing total cost by 13.1 and risk score by 0.087, confirming that closed-loop verification is the most critical factor for mitigating decision-induced risks.

Convergence Analysis:
The VARL converged efficiently with an average of 1.53 iterations per decision cycle. For structured disruptions (natural disasters), 75.8% of plans were accepted in the first iteration. For unstructured disruptions (geopolitical conflicts), more iterations were required, but escalation to human review occurred in less than 4.2% of cases.

5. Significance and Claims

The paper claims that MARS-Agent addresses the critical gap between the reflexive nature of cross-border disruptions and the open-loop paradigm of current methods. By integrating DT-RAG and VARL, the framework enables:

Proactive Risk Perception: Moving beyond reactive queries to detect risks triggered by both environmental changes and internal decisions.
Self-Correcting Deliberation: Transforming the decision pipeline into a closed-loop process where the system verifies that its own corrective actions do not introduce secondary vulnerabilities.
Balanced Optimization: Achieving an optimal trade-off between cost efficiency, delivery speed, and risk exposure in dynamic, heterogeneous environments.

The authors emphasize that while the computational overhead is higher than single-turn LLM methods (approx. 4.2x latency), it remains within acceptable ranges for daily supply chain decision cycles, offering a robust solution for resilient cross-border e-commerce operations.