📈 economics

Bounded Rationality and Resilience in Global Rare Earth Supply Chains: An LLM-Powered Multi-Agent Simulation

This paper introduces an LLM-powered multi-agent simulation framework that demonstrates how modeling sovereign decision-makers with bounded rationality significantly enhances the resilience and cost-efficiency of global rare earth supply chains compared to traditional static optimization models during geopolitical shocks.

Original authors: junjie liang, ligang xu, weifeng wang, mengtao song, Siming wei

Published 2026-07-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: junjie liang, ligang xu, weifeng wang, mengtao song, Siming wei

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

Imagine the global supply chain for rare earth elements (the special metals inside your phone, electric cars, and wind turbines) as a giant, high-stakes game of "keep-away" played on a board with only five major players. For years, experts tried to predict what would happen if one player suddenly slammed the door shut (like an export ban). They used old-school computer programs that acted like hyper-competitive chess masters: always calculating the single "perfect" move to win, assuming everyone else was also a perfect calculator.

But here's the twist: in the real world, countries don't act like perfect chess computers. They act like humans—sometimes panicking, sometimes hoarding, and often just trying to get "good enough" results without burning out.

This paper introduces a new way to simulate these games. Instead of using rigid chess computers, the researchers built a digital world where the players are powered by Large Language Models (LLMs)—the same kind of AI that writes essays and chats with you. These AI agents are programmed to be "boundedly rational," meaning they act like real humans: they have limited information, they get tired, and they settle for a solution that works well enough rather than the mathematically perfect one.

The Big Surprise: "Perfect" Players Lose

The researchers ran thousands of simulations to see who would keep the rare earth metals flowing during a crisis. They tested four types of players:

The Random Player: Just picks moves at random.
The Rule-Follower: Follows a strict, pre-written checklist.
The Greedy Optimizer: The "perfect" chess master that tries to fix every tiny problem immediately.
The LLM Agent: The human-like AI that thinks, adapts, and sometimes says, "Eh, it's fine for now."

The main finding is a bit counterintuitive. When the pressure was low (a calm market), the Greedy Optimizers went crazy. They tried to fix problems that didn't really exist, constantly rearranging trade routes and spending huge amounts of money on changes that weren't needed. The paper calls this "structural over-reactivity." It's like a firefighter who floods an entire house because they saw a single spark in the kitchen; they "saved" the house from fire, but they ruined it with water.

In contrast, the LLM agents acted more like sensible adults. They looked at the situation, realized the stress wasn't that bad, and decided not to waste energy on massive changes. They kept their costs low and their supply chains stable. The researchers measured this using a new score called the Cost Efficiency Ratio (CER). In calm times, the LLM agents had a high, positive score (efficient), while the Greedy Optimizers had a negative score (wasteful).

When Things Get Really Bad

However, the paper notes that when the stress was extreme (a total blockage of trade), the Greedy Optimizers actually did a decent job because the situation was so dire that any aggressive action was better than doing nothing. But the LLM agents still held their own, adapting quickly to find "good enough" solutions without panicking.

Interestingly, the Random Player sometimes managed to keep the supply gap low just by luck, but this came at a terrible cost. The paper explains that the Random player was like someone who throws a million darts at a board; eventually, they hit the bullseye, but they wasted a million darts to do it. The LLM agents, however, hit the target with precision and without the waste.

What This Means for the Real World

The paper argues that the old way of thinking—trying to build a supply chain that is perfectly optimized for every possible disaster—is actually dangerous. It makes countries and companies too jumpy. Instead, the simulations suggest that bounded rationality (acting with human-like limits and common sense) is actually a superpower for resilience.

The researchers found that the LLM agents naturally preferred finding new friends to trade with (diversifying) rather than just hiding metals in a basement (hoarding). This matches what actually happened during the real 2010 rare earth crisis, where countries like Japan scrambled to find new suppliers rather than just stockpiling.

How Sure Are We?

It's important to remember that these results come from computer simulations, not real-world history books. The researchers built a digital model based on trade data from 2011, 2018, and 2023, and they ran it 1,008 times with different scenarios. While the results are very consistent across different AI models (like Qwen and GLM), the paper admits that the AI might be "remembering" facts about the 2010 crisis from its training data, rather than purely figuring it out from scratch.

So, while the paper doesn't claim to have "solved" global trade, it strongly suggests that treating countries like flexible, human-like thinkers (using LLMs) gives us a much better picture of how the world will actually react to a crisis than treating them like perfect, unfeeling robots. The lesson? In a chaotic world, being "good enough" and adaptable is often better than trying to be perfect.

Technical Summary: Bounded Rationality and Resilience in Global Rare Earth Supply Chains

1. Problem Statement

Global Rare Earth Element (REE) supply chains are structurally vulnerable due to extreme geographic concentration in processing and trade, making them highly susceptible to geopolitical shocks such as export controls. Conventional methods for evaluating supply chain resilience—specifically static network models and local optimization algorithms (e.g., linear programming, mixed-integer programming)—fail to capture the complex behavioral dynamics driving multi-constraint path disruptions. These traditional approaches rely on assumptions of perfect information and absolute economic rationality, rendering them unable to internalize non-equilibrium phenomena such as strategic hoarding, panic amplification, and cognitive heterogeneity during crises. Furthermore, existing agent-based models often treat sovereign states as passive conduits or rely on rigid, pre-programmed decision rules that prevent the discovery of emergent strategies like gray-market transshipment or alliance activation.

2. Methodology

The authors propose a novel multi-agent simulation framework that integrates Large Language Models (LLMs) to instantiate bounded rationality within sovereign decision-makers. The framework operates on an empirically calibrated REE trade network and employs a unique Delta-Buffer synchronous market-clearing mechanism.

2.1 Physical Network Construction

Data Source: UN Comtrade database (2011, 2018, 2023).
Topology: A 5-node weighted directed graph representing major producers and consumers: China (CHN), United States (USA), Japan (JPN), Myanmar (MMR), and Malaysia (MYS).
Calibration: Trade flows are deflated to 2010 constant dollars. Initial inventory is set at 30% of annual demand (based on IEA benchmarks).
Metrics: Betweenness centrality is calculated using effective distance ( $d_{ij} = 1/w_{ij}$ ) to reflect trade volume intensity.

2.2 Agent Architecture

The simulation compares four distinct agent architectures:

Random: Uniform random action selection (theoretical lower bound).
Rule-based: Deterministic decision trees based on supply gap thresholds.
Greedy: Full-information myopic local optimizer that minimizes immediate supply gaps without anticipating cascading effects.
LLM: Boundedly rational agents using frontier models (DeepSeek-Chat V3, Qwen-Plus, GLM-4-Flash) at a temperature of 0.5. These agents receive structured state vectors (supply gap, inventory status, industry risk, alliances, price multipliers, and past actions) and generate context-conditioned strategic responses without ex-ante optimization functions.

2.3 Key Mechanisms

Delta-Buffer Synchronous Market Clearing: To eliminate sequential move advantages and information asymmetry, all agents observe the global state simultaneously, select actions independently, and submit "delta" vectors (changes to trade flows and inventory). These deltas are accumulated in a buffer and applied atomically to the network in a single settlement phase. This ensures mass conservation and mimics the "closed door" nature of sovereign policy formulation.
Demand–Inventory Decoupling: The framework strictly separates base demand from inventory actions. Hoarding reroutes production to storage without inflating the base demand signal, preventing "phantom demand" artifacts that could destabilize the simulation.
Action Space: Agents choose from six discrete actions, including searching for alternative sources, hoarding inventory, seeking ally support, restricting exports, gray transshipment, and releasing inventory.

2.4 Experimental Design

A full-factorial $3 \times 3 \times 4$ design was employed:

Years: 2011 (high stress), 2018 (post-crisis normalization), 2023 (current stress).
Shock Intensity: 15%, 30%, and 50% export control reductions on China.
Agent Models: Random, Rule-based, Greedy, and LLM.
Replication: 28 Monte Carlo runs per configuration (Total: 1,008 runs).

3. Key Contributions

The paper makes three primary contributions:

Methodological (Agentic LLM Simulation): It extends LLM architectures to sovereign strategic interactions within a physical trade network. Unlike baseline social simulacra, these agents are embedded in an empirically calibrated, multi-constraint context, allowing for the isolation of genuine in-context reasoning and emergent strategies (e.g., gray-market routing) without explicit static rule sets.
Mechanistic (CER-Resilience Coupling): The authors identify a fundamental regime shift in network resilience governed by the interaction between cognitive orientation and stress levels. They introduce the Cost Efficiency Ratio (CER), defined as the log-ratio of supply retention to restructuring expenditure. Simulations reveal that deterministic optimizing agents exhibit severe structural over-reactivity under low-stress conditions (negative CER), whereas LLM-powered agents dynamically calibrate efforts to maintain positive efficiency, computationally substantiating Simon's satisficing paradigm.
Metric (Contextualized CER): The paper formalizes the CER metric to endogenously capture the trade-off between material availability and path-restructuring costs. This addresses the limitation of traditional metrics that rely on marginal gap minimization, which can mask systemic inefficiencies caused by hyper-aggressive rerouting.

4. Results

4.1 Behavioral Diversity and Entropy

LLM agents exhibit a "bounded-entropy" signature. Their policy entropy (2.22–2.31) occupies a meaningful intermediate range between the rigid determinism of Rule-based agents ( $H \approx 1.66$ ) and the pure randomness of Random agents ( $H \approx 2.7$ ). LLM action frequencies shift coherently toward cooperative strategies under high stress and passive management under low stress, reflecting context-sensitive satisficing.

4.2 Historical Alignment

In an exploratory directional check against the 2010 rare earth crisis, LLM agents demonstrated the strongest qualitative agreement with historical policy rankings. Specifically, they correctly prioritized "searching for alternative sources" ( $\rho = 0.7$ ) and avoided excessive "hoarding" ( $\rho = -0.6$ ), aligning with documented responses from Japan and other importers. Deterministic agents failed to capture the diplomatic dimension of crisis response.

4.3 Supply Resilience and Gap Rates

High Stress (2011): All models exhibited supply gaps. Surprisingly, the Random agent achieved the lowest aggregate gap (0.025) due to exhaustive, unbiased exploration of the action space (Ashby's Law of Requisite Variety), though at the cost of massive inefficiency. The Greedy agent performed worst (0.073) due to myopic optimization causing cascading failures.
Low Stress (2023): LLM and Rule-based agents achieved near-zero gaps. However, the Greedy agent remained elevated (0.027), indicating that rigid optimization fails to adapt to a mature, redundant network topology.

4.4 The Stress-Regime Reversal (CER)

The central finding is a qualitative regime reversal in the Cost Efficiency Ratio:

High Stress (2011): All architectures achieved positive CER; aggressive restructuring was justified.
Low Stress (2023): Greedy and Rule-based agents collapsed into negative CER, indicating structural over-reactivity. They executed costly path rerouting despite near-adequate baseline conditions.
LLM Performance: LLM agents maintained positive CER across all regimes. By dynamically calibrating restructuring efforts to the geopolitical environment, they avoided over-investment, demonstrating that bounded rationality yields more adaptive, cost-efficient resilience than rigid optimization.

4.5 Robustness

The satisficing pattern (positive CER) was robust across different LLM models (DeepSeek, Qwen, GLM), confirming that this behavior is a class-level property of frontier-scale aligned LLMs rather than a model-specific artifact.

5. Significance and Claims

The paper claims to reframe bounded rationality not as a deficiency to be corrected, but as a systemic mechanism for adaptive, cost-efficient network optimization.

Theoretical Shift: The study challenges the dominance of rigid deterministic optimization in economic theory, demonstrating that "satisficing" algorithms provide superior resilience in complex, multi-constraint environments.
Policy Implications: The findings suggest that "de-risking" and "friend-shoring" (diversified, alliance-based routing) are economically superior to unilateral "onshoring" or "over-decoupling," which the simulations show leads to systemic inefficiencies and over-reactivity.
Methodological Advance: The introduction of the CER metric provides a tool for stress-testing trade architectures, moving beyond simple gap analysis to evaluate the cost of resilience.
Limitations: The authors modestly acknowledge that the 5-node abstraction limits topological complexity and that the LLMs' pre-training on historical data may partially influence their behavioral alignment with the 2010 crisis. They call for future work to test smaller, open-weight architectures and expand the network to include intermediate processing hubs.

Ultimately, the paper argues that in an era of intensifying critical-mineral competition, deploying adaptive, boundedly rational algorithms is paramount for securing global industrial infrastructure, offering a robust critique of rigid domestic self-sufficiency strategies.