Requested, Executed and Recoverable Credit: Shared Action Ordering among LLM Borrowers in a Simulated Lending Market

Paper: arXiv 2610.15973

Authors: Victoria Ruojie Li

Abstract

AI agents built on different language models can still share the software that presents their choices. We study one shared component, the order in which a borrowing agent’s actions are listed, in a simulated collateralized lending market populated by three open-weight model families. In fixed decision states, moving the borrowing action from last to first raises valid borrowing requests by 27 percentage points. In 12 matched pairs of 15-agent markets, all agents in one market see the same order; the other spreads the six possible orders across agents. A common order moves whole markets: the four markets that listed borrowing first raised net debt by 1.8 to 2.7 times their initial unused borrowing capacity, while those that listed debt reduction first cut debt. These groups account for 96 percent of the sample variation among common markets, whose standard deviation is three times the balanced one. The borrowing cap stops 40 percent of valid borrowing requests made when borrowing is listed first, but it values collateral at the pool price, so correlated purchases raise borrowing capacity and leave positions above 70 percent of reference value. Under a 30 percent shock, an 8 percent liquidation bonus produces write-offs on loans that are fully collateralized before liquidation, and liquidation triggers that read the pool price add exposure. Common ordering does not put more borrowers into the write-off zone; it concentrates them in the markets whose shared order favored borrowing. Collateral valuation and liquidation rules determine how much of it creditors fail to recover.

Complexity vs Empirical Score

• Math Complexity: 4.0/10
• Empirical Rigor: 8.0/10
• Quadrant: Street Traders — practical and empirical, lighter on theory

Why this score: This paper presents a highly novel and empirically rigorous study on the impact of action ordering on LLM agents in a simulated financial market. The methodology is well-defined, and the results are clearly presented, making a significant contribution to understanding AI agent behavior and financial stability. The math is not overly complex, but the simulation design and statistical analysis are robust.

Research Flowchart

  flowchart TD
    A[Research Goal: Impact of Action Ordering on LLM Borrowers] --> B{Methodology: Simulated Collateralized Lending Market};
    B --> C[Inputs: LLM Agents (3 Families), Fixed Decision States, Action Orderings (6)];
    C --> D{Computational Process: Market Simulations, Action Order Manipulation};
    D --> E[Key Finding 1: Borrowing Action First -> +27pp Valid Borrowing Requests];
    E --> F[Key Finding 2: Common Order Drives Market Behavior (Borrowing vs. Debt Reduction)];
    F --> G[Key Finding 3: Collateral Valuation/Liquidation Rules Impact Recovery, Not Common Ordering of Write-offs];

Cited References (37)

1. Algorithmic monoculture and social welfare (2021) — link
2. Picking on the same person: Does algorithmic monoculture lead to outcome homogenization? (2022)
3. Large language models are not robust multiple choice selectors (2024) — link
4. Large language models sensitivity to the order of options in multiple-choice questions (2024) — link
5. Quantifying language models’ sensitivity to spurious features in prompt design or: How I learned to start worrying about prompt formatting (2024)
6. Artificial intelligence and systemic risk (2022) — link
7. The Financial Stability Implications of Artificial Intelligence (2024)
8. Deep Learning and Financial Stability (2020) — link

0 of 37 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2518 linked papers.