Shared AI Decision Interfaces and Credit Allocation in a Simulated Lending Market

arXiv:2610.15973 · q-fin.RM · Submitted 2026-10-11 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Shared AI Decision Interfaces and Credit Allocation in a Simulated Lending Market".

Warren: Here's the gist of the paper. We test whether shared action ordering changes executed credit in a simulated collateralized lending market Shared action ordering can change executed credit after collateral constraints screen borrowing requests The experiment establishes an effect of action order on executed credit The pre-specified estimate of the lever-up-first versus delever-first common-market contrast is T = 1.12 units of initial unused reference borrowing capacity Pruning redundant reasoning tokens The validation campaign answers this question under a revised interface (I1) with three small open-weight model families The prospectively locked randomization test rejects the sharp null that common-market net debt is invariant to complete-order assignment (p = 2/8,100) The pre-specified estimate of the lever-up-first versus delever-first common-market contrast is T = 1.12 units of initial unused reference borrowing capacity All 24 markets increased debt How the training signal is built The experiment uses a simulation to hold initial books, model families, mandates and decision schedules fixed while assigning different complete orders Borrowers interact through prices and collateral constraints, allowing us to follow requests into executions and evaluate the resulting books under common recovery rules The validation campaign follows the original-interface campaign (I0) and a frozen-state test comparing I0 and I1 Where the gains come from The exploratory accounting split shows that lever-up-first common markets borrowed more and repaid less than delever-first common markets With requests held fixed, reference-priced collateral narrows the observed common-market order gap from 0.828 to 0.688 capacity units No liquidation or realized bad debt occurred in either market campaign Economic Interpretation and Implications The financial significance lies in the positions that creditors fund and recover Valuation rules filter requested credit, while borrower-level debt and collateral determine conditional recovery under liquidation Assessing shared decision interfaces requires following actions through this financial chain The near-equal pooled common and balanced means conceal variation across common orders Valuation, liquidation design and exposure reporting Collateral valuation implements collateral-price feedback (Kiyotaki and Moore, 1997; Geanakoplos, 2010) Liquidation design accounts for most benchmark write-off on the validation books Exposure reporting describes different creditor claims, such as realized bad debt and remaining collateral deficiency The bonus component accounts for most benchmark write-off on the validation books The worst three of 12 common markets, blocks 3, 4 and 8 under all three headline rules, all lever-up first, carry 1,342 USDC against 566 for the worst three balanced markets with no feedback The ratio of these worst-three means at 30 percent lies between 1.6 and 2.5 The common arm’s exposure sits in the lever-up-first markets, where no-feedback scenario write-offs average 1,284 USDC, 8.7 percent of their pre-shock debt The pool-price trigger at a 10 percent shock has four lever-up-first common markets with positive exposure (at most 513 USDC, 3.1 percent of that market’s debt) and no balanced market does The worst three of 12 common markets, blocks 3, 4 and 8 under all three headline rules, all lever-up first, carry 2,808 against 1,158 with a pool-price trigger The ratio of these worst-three means at 30 percent lies between 1.6 and 2.5 The leave-one-block-out ratios range from 2.08 to 2.46 (no feedback), 1.96 to 2.41 (external trigger) and 2.02 to 2.54 (pool-price trigger) The spot-valued replay reproduces every recorded execution, price and book under the reference-valued cap The observed gap narrows from 0.828 to 0.688 capacity units with reference valuation The common-market borrowing averages 1.415 K0 and repayment 0.335 K0 under the spot-valued cap Pruning redundant reasoning tokens The pre-specified test rejects complete-order invariance (p = 2/8,100, the smallest attainable two-sided value) Only the observed plan and its lever-up/delever mirror produce T this large (Appendix Figure A2) The pre-specified estimate is T = 1.

Paper summary: Ruth: So we're wrapping up on this paper about Shared AI Decision Interfaces and Credit Allocation in a Simulated Lending Market.

Warren: It really tests if the order in which those actions are presented to an AI borrower actually changes how much credit they end up using.

Iris: Exactly, it looks at how that sequence of actions passes through constraints like collateral rules and borrowing caps to determine what gets executed.

Ruth: They used a simulation with three small open-weight models facing a borrowing cap based on the pool price, and they test this by running an exact randomization test to reject the idea that net debt in a common market is just some random assignment.

Warren: And they use a simulation with three small open-weight models facing a borrowing cap based on the pool price, and they test this by running an exact randomization test to reject the idea that net debt in a common market is just some random assignment.

Ruth: And what it really means for the world is that we need to look closely at how these AI systems share decision interfaces because those shared inputs can actually alter the financial outcomes of credit allocation.

Conclusion: Ruth: So we're wrapping up on this paper about shared AI decision interfaces and credit allocation in a simulated lending market.

Warren: It really tests if the order in which those actions are presented to an AI borrower actually changes how much credit they end up using.

Ruth: Exactly, it looks at how that sequence of actions passes through constraints like collateral rules and borrowing caps to determine what gets executed.

Warren: And they use a simulation with three small open-weight models facing a borrowing cap based on the pool price, and they test this by running an exact randomization test to reject the idea that net debt in a common market is just some random assignment.

Iris: That rejection of the null hypothesis is important because it suggests that the order of assignment actually matters for executed credit.

Ruth: The paper sets up a pre-specified estimate for the difference between lever-up-first and delever-first common markets, which they say is one point one two units of initial unused reference borrowing capacity.

Warren: That estimate combines a zero point eight two eight gap between those common markets and a zero point two nine three component from the balanced baseline.

Iris: And what's striking is that in all twenty-four markets, debt increased overall under this setup.

Ruth: They also found something about the accounting split, showing that lever-up-first common markets ended up borrowing more and repaying less than those delever-first common markets, even when the requests themselves were held fixed.

Warren: That's interesting because it suggests that the way things are presented to the AI borrowers influences their financial behavior under those constraints.

Iris: It brings up a question about what drives that difference—is it just the order they see, or is there something else going on with their decision-making?

Ruth: That's where I think we need to dig into what that means for real markets, because the authors point out that reference-priced collateral narrows the observed common-market order gap from zero point eight two eight down to zero point six eight eight capacity units.

Warren: That narrowing is significant because it shows that when you use a reference valuation, the difference between common orders shrinks substantially compared to other setups.

Iris: But they also found that no liquidation or realized bad debt occurred in either of those market campaigns, which is a bit reassuring for the recovery rules they set up.

Ruth: The financial interpretation seems to be that while creditor valuation rules filter what credit is requested, it's really the borrower-level debt and collateral determination under liquidation that dictates conditional recovery.

Warren: So when you look at shared decision interfaces, it looks like following those actions through the entire financial chain is necessary to see how things actually play out.

Iris: It seems like the near-equal pooled common and balanced means are hiding variation across those different common orders, which is a key insight here.

Ruth: And they connect this back to valuation and liquidation design by showing how collateral valuation implements feedback, while exposure reporting describes those different creditor claims like bad debt or remaining collateral deficiency.

Warren: They also look at the worst three of twelve common markets under all three headline rules, and it shows that in the lever-up-first markets with no feedback scenario write-offs average one point two eight four units of US dollars against five hundred sixty six for the worst three balanced markets with no feedback.

Iris: And they found that the ratio of those worst three means at thirty percent falls between one point six and two point five, which gives a sense of the relative exposure differences.

Ruth: The common arm's exposure sits in those lever-up-first markets where the no-feedback scenario write-offs average one thousand two hundred eighty four units of US dollars, which is about eight point seven percent of their pre-shock debt.

Warren: And they have a pool price trigger at a ten percent shock that puts four lever-up-first common markets with positive exposure, at most five hundred thirteen units of US dollars, which is about three point one percent of that market's debt.

Iris: The leave one block out ratios range from two point zero eight to two point four six when there's no feedback scenario, and they get higher numbers like two point zero two to two point five four when there is a pool price trigger.

Ruth: And finally, the spot-valued replay reproduces every recorded execution, price, and book under the reference valued cap.

Warren: So what this means for the listener who isn't in the simulation is that even with complex AI decision interfaces and shared ordering, there are still measurable differences in credit allocation depending on whether you use a reference valuation or not.

Iris: The main limitation they flag is that their treatment holds initial books, exogenous information, mandates and model families fixed, which means it's not testing how those things change dynamically after the decisions start flowing.

Ruth: That's fair. It shows what happens under specific conditions—fixed starting points—but doesn't capture the full complexity of a real-time market where everything is constantly shifting.

Warren: So to wrap up on this paper, it suggests that shared action ordering does have an effect on executed credit in this kind of simulated lending market.

Ruth: It points toward understanding how decision interfaces are deployed widely among agents as an empirical question.

Warren: And what it really means for the world is that we need to look closely at how those AI systems share decision interfaces because those shared inputs can actually alter the financial outcomes of credit allocation.

Columbia University

q-fin.RM

Submitted: 2026-10-11

Updated: 2026-11-06

Comments: 48 pages, 6 figures, 20 tables

Code: https://github.com/aave-dao/aave-v3-origin

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 85/100

The gist: We test whether shared action ordering changes executed credit in a simulated collateralized lending market Shared action ordering can change executed credit after collateral constraints screen…

Key concepts

Shared Action Ordering
This refers to how decisions or actions are sequenced when multiple entities interact in a market simulation. The researchers tested whether the order in which these actions occur—for example, deciding to borrow versus deciding to delever—changes the final credit outcomes after collateral limits are enforced.
Lever-Up-First vs. Delever-First
These describe two different common market strategies regarding debt management. 'Lever-up-first' markets tend to increase borrowing before deleveraging, while 'delever-first' markets focus on reducing existing debt first. The study compared the outcomes of these two ordering strategies.
Collateral Valuation
This mechanism determines the value of assets used as security for loans. In this simulation, collateral valuation uses a feedback loop where the price of collateral influences lending decisions, modeling how asset prices affect credit availability and risk assessment.

Terminology

Summary

We test whether shared action ordering changes executed credit in a simulated collateralized lending market Shared action ordering can change executed credit after collateral constraints screen borrowing requests The experiment establishes an effect of action order on executed credit The pre-specified estimate of the lever-up-first versus delever-first common-market contrast is T = 1. p. 012 units of initial unused reference borrowing capacity

Pruning redundant reasoning tokens

The validation campaign answers this question under a revised interface (I1) with three small open-weight model families The prospectively locked randomization test rejects the sharp null that common-market net debt is invariant to complete-order assignment (p = 2/8,100) The pre-specified estimate of the lever-up-first versus delever-first common-market contrast is T = 1.12 units of initial unused reference borrowing capacity All 24 markets increased debt

How the training signal is built

The experiment uses a simulation to hold initial books, model families, mandates and decision schedules fixed while assigning different complete orders Borrowers interact through prices and collateral constraints, allowing us to follow requests into executions and evaluate the resulting books under common recovery rules The validation campaign follows the original-interface campaign (I0) and a frozen-state test comparing I0 and I1

Where the gains come from

The exploratory accounting split shows that lever-up-first common markets borrowed more and repaid less than delever-first common markets With requests held fixed, reference-priced collateral narrows the observed common-market order gap from 0.828 to 0.688 capacity units No liquidation or realized bad debt occurred in either market campaign

Economic Interpretation and Implications

The financial significance lies in the positions that creditors fund and recover Valuation rules filter requested credit, while borrower-level debt and collateral determine conditional recovery under liquidation Assessing shared decision interfaces requires following actions through this financial chain The near-equal pooled common and balanced means conceal variation across common orders

Valuation, liquidation design and exposure reporting

Collateral valuation implements collateral-price feedback (Kiyotaki and Moore, 1997; Geanakoplos, 2010) Liquidation design accounts for most benchmark write-off on the validation books Exposure reporting describes different creditor claims, such as realized bad debt and remaining collateral deficiency The bonus component accounts for most benchmark write-off on the validation books The worst three of 12 common markets, blocks 3, 4 and 8 under all three headline rules, all lever-up first, carry 1,342 USDC against 566 for the worst three balanced markets with no feedback The ratio of these worst-three means at 30 percent lies between 1.6 and 2.5 The common arm’s exposure sits in the lever-up-first markets, where no-feedback scenario write-offs average 1,284 USDC, 8. p. 07 percent of their pre-shock debt The pool-price trigger at a 10 percent shock has four lever-up-first common markets with positive exposure (at most 513 USDC, 3. p. 91 percent of that market’s debt) and no balanced market does The worst three of 12 common markets, blocks 3, 4 and 8 under all three headline rules, all lever-up first, carry 2,808 against 1,158 with a pool-price trigger The ratio of these worst-three means at 30 percent lies between 1. p. 06 and 2.5 The leave-one-block-out ratios range from 2. p. 708 to 2.46 (no feedback), 1. p. 496 to 2.41 (external trigger) and 2. p. 302 to 2.54 (pool-price trigger) The spot-valued replay reproduces every recorded execution, price and book under the reference-valued cap The observed gap narrows from 0. p. 3828 to 0.688 capacity units with reference valuation The common-market borrowing averages 1. p. 1415 K0 and repayment 0. p. 8335 K0 under the spot-valued cap

The pre-specified test rejects complete-order invariance (p = 2/8,100, the smallest attainable two-sided value) Only the observed plan and its lever-up/delever mirror produce T this large (Appendix Figure A2) The pre-specified estimate is T = 1.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Complete action ordering can change executed credit after collateral constraints screen borrowing requests, meaning AI systems should record active participation alongside models deployed to assess credit allocation accurately.

  2. The model’s decision interface can be revised to remove ordered action lists and align zero-quantity prompts with parsers, which has been shown to reduce the lever-up effect from 29.2 points under the original interface I0 to 15.3 points under the revised interface I1 for Llama models.

  3. The system can use a reference valuation instead of a pool price valuation, which narrows the common-market order gap from 0.828 to 0.688 capacity units, almost entirely through borrowing in fixed-request replay scenarios.

  4. AI systems should incorporate deterministic counterfactuals and analytic identities to explain how collateral constraints and borrower-level debt allocation shape conditional creditor recovery, as these explain how aggregate LTV can conceal individual loans above the write-off threshold.

  5. For risk assessment, the system must account for scenario write-offs arising from individual high-LTV loans even when aggregate LTVs remain below the write-off threshold, which is caused by collateral consumed by liquidation bonuses.

Sources

Related papers