acceptodds
Under review as a conference paper at ICLR 2027

Financial Fragility in Societies of LLM Agents: Coordination Failures and Stabilizing Mechanisms

Abstract

Individually protective decisions can produce avoidable collective failures. As large language model (LLM) agents take on greater roles in financial decision-making, financial AI safety must therefore be considered not only at the level of individual agents, but also at the level of the systems they jointly create. We study this problem with FRAIL, a controlled experimental framework that places LLM agents in three dynamic financial environments—bank runs, debt rollover, and reward crowdfunding—where agents' decisions reshape the financial conditions faced by others. Across seven leading LLMs, we find widespread collective fragility even when no agent is instructed to destabilize the system: 77% of baseline bank-run episodes and 83% of debt-rollover episodes end in failure. We then compare three interaction mechanisms based on compensated commitments, centralized commitment agreements, and participant-led coalitions. All three improve aggregate outcomes, but no single mechanism performs best across all financial structures. Across mechanisms, successful stabilization shares a common temporal pattern: broad commitment forms early, before defensive behavior becomes self-reinforcing. Our findings show that individually capable agents do not automatically form safe financial systems, highlighting system-level evaluation and interaction design as central problems for financial AI safety. Code is available at https://anonymous.4open.science/r/FinFrail-CF26.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agents, Multi-Agent Systems, Financial AI Safety, Financial Fragility, Coordination Failure, Mechanism Design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_fdNBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The authors analyze MAS of LLM agents acting within a financial institution and observe possible failures that can emerge from individual rational outcomes, such as trying to protect themselves. They develop a framework, called FRAIL, to analyze the fragility of isolated economic institutions populated by autonomous LLM agents. They show that such failures can happen and propose three mechanisms to alleviate such problems.

Critical strengths

The authors study a very interesting new cooperation problem of multi-agent LLMs within financial institutions. The setting is novel, and the experimental setting looks solid:

  • The authors develop a simulation environment focused on risk in a financial environment, with dynamic financial trajectories that other related works have not explored in depth.
  • The experimental analysis of the paper is strong; the authors show that failure exists, show that the chosen mechanism helps in alleviating the coordination problems and analyze reasons why such mechanisms help. They also ablate the environments with favorable or tighter conditions. They offer confidence bounds and different seed runs.
  • The article is well written and clear. The images are clear and aid in understanding the setting. I also liked that the paper provided a glossary for CS people who might not be familiar with the finance background. These are small signals of attention to detail during the writing of the paper.
Critical weaknesses

I have a strong methodological issue regarding the parsing of the choice of actions in the experiments. The paper reports that it parses a “defensive intent” or a probability of choosing one or the other, and then samples using this probability. Yet, I argue that doesn’t correspond to correct parsing of the model's choice. In such discrete action models, the models in reality choose a single action. The fact that you then sample is something only your experimental setting does, but LLMs out there do clear actions instead. For example, if a model gives a 10% probability of withdrawal, having 12 agents means we expect at least one agent to withdraw even if all might give a 10% withdrawal possibility. If this were the case, then they would instead all choose not to withdraw. This would then feed into the next round, creating a bad feedback cycle.
I would like to see evidence with code of this not being relevant to raise my score. I don’t believe the “This design avoids brittle dependence on exact action parsing and, because the reports are graded rather than binary, provides an agent-level measure of defensive intent;” since modern models are not brittle anymore on action parsing if you prompt correctly.

Since your work is mostly experimental, evaluating the soundness of your code and setting is crucial for good judgment. I read the logs in your anonymous code, and indeed models choose 5 or 10 probability even when their reasoning is No redemptions occurred last round, and the institution appears comfortably solvent relative to promised obligations. I prefer keeping the deposit to earn the 1.114 maturity payment, with only a small withdrawal probability for unforeseen risk.

Another Issue, is that while authors offer ways to solve the coordination problem that they find within the financial systems, they offer no discussion on how their findings might offer insight into debugging or engineering some multi-agent system that can be alleviated by those issues. Especially in finance, where bad engineering can cause billions of dollars lost, we might care more about principled engineering of these systems instead of testing some mechanisms and empirically seeing that they make the problem less severe. The paper does the latter. This connects to my first question later when I ask the authors to explain how they came up with the mechanisms and the financial environments.

The model is still a highly abstract and formal model of how the specific institution there works; for example, the statement that “The institution fails immediately if it cannot pay a redemption request during processing,” is ok for the experiment, but might not be true of real financial institutions that the work is attempting to model. Real institutions may find other devices to finance themselves, especially if it's a state bank. But I agree that this might be out of scope for a single paper. Another drawback is that as a society we are already aware of the risks that automated trading/financial agents can have, e.g., Flash Crash. So, in my opinion, the risk that the paper is studying is quite limited. A more interesting setting could be blockchain or other newer financial institutions where regulation is less stringent.

I regard the last two issues as main reasons why I don’t lean towards a clear acceptance score even if the main methodological issue is resolved. I think the paper is well positioned, and it identified a true problem with LLMMAS with a couple of possible mechanisms to alleviate the problem, which makes me willing to vouch for a weak acceptance given resolution of the main methodological concern.

In addition, looking at the prompts, I see words such as [GAME PARAMETERS], which might tell the model that the environment is not “real”, which might produce some eval awareness problems. But since the model itself is abstract, as explained before, I don’t see this as a big issue in this evaluation scenario.

Single runs also all use the same model, even if models might differ between runs. I think this is a minor issue.
Another thing is that since you have seeded the runs, and the seed generates the original deposit (from the code), every model, mechanism and regime has the same amount deposited. This is also a minor issue, but might be helpful if you also ablate the amount deposited and see how much that influences the bank run.

Questions for authors
  • I am curious to know why and how the authors selected the three experimental financial scenarios to test. There are also other cases where we can see such institutions failing. Finance is the modern version of some more classical cooperation dilemmas. In the same manner, I would be interested in how the authors chose the mechanisms.
  • I am confused by “economic regime” appearing for the first time in section 4.3. What is it referring to?
  • For example, I found [INTERPRETATION REMINDERS] - Heavy redemptions in previous rounds may indicate elevated risk. in the prompt. Together with my methodological issue regarding probabilities, I believe the experiment may have mistaken results and bias the systems towards higher failures than expected. Can you share the logs of the Mechanisms too? With multiple rounds.
  • What share of exits came from reports below about 20%?
11 comments
AuthorsOpenReview ↗

Response to reviewer fdNB (Part1/4)

Thank you for the careful inspection and constructive comments. In response to your major concern (methodological issue about the probabilistic action choice), we completed a full-scale robustness test to check results under direct action choices. Also, we acknowledge the limitation of this paper you mentioned in the official review, and would like to further articulate our research question and its potential implication. Finally, we will provide direct answers to your specific questions.

Direct-Action Results, Execution Code, and Traces

Response: We agree that modern LLMs can return explicit actions, and have removed the action-parsing justification in the manuscript. Inspired by mixed strategies in game theory, we ask participants to report action probabilities so that we can track how strongly they favor different choices as the financial situation changes. To ensure our results do not rely on the probabilistic action choice, we reran the experiments using direct action choice.

We changed the prompts so that models output an action directly, such as withdrawing or keeping a deposit, instead of outputting a probability for the environment to sample. The execution code then follows the selected action under the original financial and commitment rules. Reviewer qbAF raised the same concern, and we treated it as a priority in our additional evaluation. The results show that collective failures persist and commitment mechanisms improve aggregate outcomes. Below we provide the results, the execution code, and the recorded action sequences.

Aggregate results. The reported comparison in Table 14 gives the following mean normalized welfare on a percentage scale:

EnvironmentM0M1M2M3
Bank run40.84%62.19%86.49%62.86%
Debt rollover11.57%42.85%43.55%32.97%
Reward crowdfunding80.79%89.25%88.75%91.87%

The comparison supports the main qualitative findings under direct action selection (Section 4.1; Appendix J.3). The table reports welfare, with 0% and 100% denoting the environment’s worst and best welfare reference points. Each cell summarizes 35 episodes, five per model; bold marks the largest mean in each row.

Execution code. The direct-action branch skips probability sampling. The following code in bankrun/sim/runner.py handles the selected action under the original commitment constraints, then continues before the probability-sampling branch:

if cfg.action_interface == "direct":
    if dep.locked:
        if out["effective_action"] == "redeem":
            state.stats["overrides"] += 1
            log.log(t, "withdrawal", dep_id, "override",
                    {"reported_action": out["action"],
                     "forced_action": "stay"})
    elif out["effective_action"] == "redeem":
        requests.append(dep_id)
    continue

Recorded trajectories. A recorded failure arises from valid discrete choices. In a GLM-5.3-Flash M0 bank-run episode, all 15 depositors initially choose stay. Their subsequent choices develop as follows:

RoundDepositors explicitly choosing withdrawalState at round end
1–60 in each roundContinues
71Continues
83Continues
97Continues
107Insolvency

These counts are depositors explicitly selecting redeem in each round; actual payments remain subject to the financial rules. Every financial response in this baseline episode is a valid action output, with zero corrections and zero repairs. The bank’s failure therefore occurs in a trajectory of explicit action choices.

In the accompanying direct-action M2 example, an agreement forms in round 1, commitments are announced, and the bank survives to maturity in round 12 without withdrawals. The direct-action walkthrough shows both episodes round by round and points to the relevant records. The additional evidence package ("FRAIL_rebuttal_evidence.zip") includes their saved prompts and responses for every recorded call, complete event streams, and metrics, together with the execution code.

AuthorsOpenReview ↗

Response to reviewer fdNB (Part2/4)

Research Contribution and Potential Engineering Implications

We humbly admit that our research has limited insight into debugging and engineering a multi-agent system. We focus on a different source of risk that is mostly parallel to the failure from engineering mistakes. The Coordination Fragility problem we studied can arise even when every agent operates correctly: individually protective withdrawals or redemptions change the decision environment faced by other agents and create system-level externalities. This is by the nature of these dilemmas and any kind of engineering advancement will not solve such a problem. Our contribution is to provide a reproducible setting in which LLM decisions alter a shared financial state, allowing us to trace collective failure and compare executable commitment arrangements.

We agree that the risks of automated financial activity are already recognized. Our unique contribution is to characterize coordination fragility more precisely. Our experiments deepen understanding of these dynamics and evaluate mechanisms that can mitigate them. Also, our abstract scenarios are not restricted to traditional institutions or particular regulatory regimes. We focus on shared dilemmas and coordination mechanisms rather than any specific institutional form, thus our abstract scenarios are inclusive of coordination challenges relevant to blockchain-based and other emerging financial arrangements.

Although detailed engineering guidance is not our objective, our findings may offer some useful design implications as well. For example, systems should provide authenticated communication channels and make commitments machine-verifiable from the start, helping agents coordinate more reliably. These are practical considerations for building more reliable multi-agent systems.

Responses to Your Questions

Q1: Why were these financial scenarios and mechanisms selected?

A1. Given the problem we try to tackle is the coordination fragility in multi-agent systems, we find financial decision scenarios most fit to such a topic and have the most practical implications. The principles we followed when we designed the experimental games are simple and clear: 1. The games must be representative of a family of real-world scenarios where coordination problem is an important issue and AI agents will possibly participate with autonomy in the foreseeable future. 2. The scenarios should be common, important and not difficult to understand. 3. The games should each capture distinct structures from real-world scenarios, depict different forms of coordination problems.

In a similar manner, we have chosen the 3 mechanisms that satisfy: 1. The mechanisms must resemble real-world mechanisms designed to mitigate the coordination problem. 2. The mechanisms must be distinct from each other in terms of how they affect action choices of the participants. 3. The mechanisms can be abstracted in a relatively simple fashion that enables the simulation results to be interpretable. The final chosen combination of scenarios and mechanisms follows all the principles above and together covers a broad scope of our research question.

The selected scenarios represent three distinct funding tasks: retaining existing funds in bank runs, renewing maturing funding in debt rollover, and assembling contributions to meet a goal in crowdfunding. Their different timing and funding needs allow us to study coordination across several financial structures. And the mechanisms cover compensated retention offers, centrally organized unpaid agreements, and participant-led coalitions, which each stands for a group of collective action protocols.

Q2: What does “economic regime” refer to?

A2. Thank you for pointing out this ambiguity. Here, an economic regime means a set of financial parameters. We create more favorable or tighter conditions by changing quantities such as asset returns, interest rates, available cash, and funding requirements. For example, the more favorable bank-run setting increases initial cash and asset returns. These are numerical changes to the environment. Section 4.3 describes the changes directly, Appendix D defines the term at its first appearance, and Table 6 in Appendix D.2 lists the parameter values.

AuthorsOpenReview ↗

Response to reviewer fdNB (Part3/4)

Q3: How do you address the prompt wording concern, and can you share multi-round mechanism logs?

A3. We agree that prompt wording is important to examine. We clarify how the reminder is used and then present the wording checks.

Risk reminder. The reminder is static and appears in withdrawal-decision prompts under M0–M3, including the first round when previous redemptions are zero. All four conditions also report the actual fraction redeemed in the previous round. The reminder describes a financial feedback present in the environment: redemptions can deplete cash and require discounted asset sales.

Rephrasing and economic descriptions. We ran 540 episodes using either a rewording that preserves the original instructions' meaning or an added optimistic or pessimistic economic description intended to nudge the agents' outlook. These tests change the text shown to the models while keeping the financial parameters and mechanism rules fixed. They also address reviewer qbAF's question about prompt variations.

Removing game and simulation wording. Another 180 episodes replaced fixed references to games and simulations with neutral financial descriptions. For example, [GAME PARAMETERS] becomes [FINANCIAL PARAMETERS], and “You are simulating the behavior of a single depositor” becomes “You are one depositor.” Financial rules and the risk reminder are retained.

After this wording change, M0 survival is 6.67% in bank runs and 0.00% in debt rollover, while commitment mechanisms continue to improve outcomes (Section 4.3; Appendices J.5–J.6). For your request for multi-round mechanism logs, the reading index (FRAIL_rebuttal_evidence/reading_index.md) lists 24 original-experiment episodes covering M0–M3 in all three environments for GLM and GPT-terra. Each includes the complete event stream for all participants, plus separately saved prompts and responses for participant 01. The two direct-action examples described above additionally include saved prompts and responses for every recorded call.

Q4: What share of exits came from reports below about 20%?

A4. We checked this directly in the recorded withdrawals. We matched each paid withdrawal to the same depositor’s reported probability in the same round across the 140 main-experiment bank-run episodes:

ScopeReported withdrawal probabilityShare of paid withdrawals
All conditionsStrictly below 20%20.09%
All conditionsAt or below 20%22.14%
M0 onlyStrictly below 20%23.86%

The denominators are 682 paid withdrawals across all conditions and 285 under M0. The paid_withdrawals_682.csv (FRAIL_rebuttal_evidence/checks/paid_withdrawals_682.csv) lists the reported probability and actual payment for the same depositor in the same round. It includes event locations and the indicators used for the percentages above. The check guide (FRAIL_rebuttal_evidence/checks/README.md) links to the supporting records and verification code. These counts describe the original experiments, in which models report probabilities and the simulator samples actions; the direct-action results above test explicit choices.

Reviewer_qbAFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces FRAIL, a framework for studying financial failure modes in societies of LLM agents. FRAIL consists of three environments grounded in classic finance models: bank runs, debt rollover, and reward crowdfunding. Across seven LLMs, they find frequent collective failure under the baseline: 77% of bank-run and 83% of debt-rollover episodes fail, and 23% of crowdfunding campaigns never launch. The authors then compare three commitment mechanisms: compensated commitments paid by a central player, centralized unpaid agreements, and participant-led coalitions. They find that all three improve welfare and success on average but the best mechanism depends on the game.

Critical strengths

The paper studies an important problem - how individual rational decisions by LLM agents can lead to collective financial failure as LLMs are increasingly used in financial decision-making processes. The proposed framework includes three games grounded in the finance literature and evaluates multiple LLMs. The framework is also dynamic, meaning each agent's decisions change the financial state and the incentives faced by other agents which mimics how one action affects others.

Critical weaknesses

M1, M2, and M3 vary along multiple axes at once, such as financial incentives, information sharing or commitment enforcement. Thus, some of the paper's conclusion such as "Communication trumps compensation," are not fully supported. It is unclear whether the improvement is due to communication, commitment, or something else. Also, M1 and M2 add an additional LLM central player that M0 and M3 lack. How much of the M1/M2 effect comes from that additional behavior rather than from the mechanism? I feel there is lot of inconsistency across M1,M2,and M3 which makes it hard to interpret findings.

The Results section focuses mainly on outcome-level metrics with limited analysis of how LLM agents made their decisions or how their reasoning changes across different scenarios. Also, the Results section is hard to follow. For instance, Sections 4.1 and 4.2 contain significant overlap and redundancy in their findings and interpretations.

Questions for authors

Have you tested whether the baseline failure rates and mechanism rankings hold under prompt variations? Have you tested whether the results remain consistent when agents directly select actions instead of reporting action probabilities?

5 comments
AuthorsOpenReview ↗

Response to reviewer qbAF (Part1/4)

Thank you for recognizing the importance of studying collective financial failures and the value of our dynamic multi-agent framework. We believe your comments on making the mechanism comparisons and behavioral analysis clearer have substantially helped us improve the quality of the paper. We will address your primary concerns first, followed by our answers to your more specific questions.

Responses to the Main Comments

W1: Inconsistency Across Mechanisms and Interpretation Difficulty

Response: We acknowledge the main issue you proposed in the review. The experimental design would appear to be inconsistent between M1, M2, and M3, but we have made our thoughtful choices of which part to include in one mechanism and not in another. In general, the key elements in the 3 mechanisms are selected as the minimal sets of elements that grant enough variation across mechanisms and embody the essence of the mechanisms we try to establish at the same time (without losing resemblance to the real world scenario). For example, let's consider the compensation mechanism (M1) in the bank run game. The initial idea is to enable the bank to keep depositors from withdrawing by offering additional reward. This will change the disclosure of the bank's financial status, because offering additional reward to a certain group of clients will worsen the liquidity constraint faced by the bank, which ought to be broadcasted to all depositors. The change in disclosure brings in another issue that the rest of the depositors (those who didn't receive the additional reward) will only interpret the bad consequence of the activated M1 commitment if they are not informed about the share of deposits locked. So the communication channel that enables the bank to send messages to its depositors about the activated commitment share is necessary as well. Such considerations appear in every case, so there is no feasible way to implement 3 distinct mechanisms and guarantee the clearest comparison at the same time.

Of course, some part of the design can be done a different way, like not adding an additional LLM central player in M1 and M2, but using a rule-based algorithm to automatically send inquiries to depositors/creditors/investors and announce commitment status. Given the project is about studying LLM agents' behavior, we chose the central agent implementation. But we are also curious about how a rule-based central player will perform in the game, so we did the following test:

(i) Replacing the central LLM with fixed rules. We made the bank's or borrower's offers, invitations, and announcements follow prespecified rules instead of LLM decisions. Participants still use LLMs to decide whether to accept or sign. For example, in bank-run M2, the rule invites all 15 depositors in round 1; each depositor decides whether to join, and the original commitment rules apply. Across 80 additional episodes, fixed-rule M1 and M2 continue to improve survival and mean welfare relative to M0. The table compares their survival rates with M0:

EnvironmentM0 referenceFixed-rule M1Fixed-rule M2
Bank run5.00%30.00%100.00%
Debt rollover20.00%55.00%80.00%

These results show that the mechanism gains remain when central decisions are made without an LLM. As long as the mechanisms can be utilized by making reasonable propositions and sharing information, the output enhancement is expected to be seen. For M2, the corresponding LLM-central-player survival rates are 70.00% in bank runs and 80.00% in debt rollover. Each percentage summarizes 20 episodes, five per model. The fixed rules and full results are reported in Section 4.3 and Appendix J.4 (Tables 15–16).

AuthorsOpenReview ↗

Response to reviewer qbAF (Part2/4)

We believe the integrity of the 3 mechanism designs is best preserved in our main results' setting, but we admit the inconsistency across mechanisms could make interpretation a challenge. So we did additional experiments to help us better understand the source of the improvement. In the ablation study, we shut down the communication channel that informs agents about the total activated commitment share, so commitments in all mechanisms only help to alleviate the fragility by locking the activated shares themselves (not by sending stabilizing signals). The details and results are as follows:

(ii) Removing information about other participants' commitments. We kept signing, fund locking, and exit procedures, but removed extra information that helps an agent learn whether others have committed to cooperate. Depending on the mechanism, this includes other participants' accepted offers, aggregate committed funds, group details, and related announcements. An agent still sees its own contract and account status, along with publicly reported financial activity such as withdrawals that have already occurred. Thus, commitments remain binding while participants have less information about one another's cooperation and exit plans. For GLM, Qwen, and DeepSeek, M3 mean welfare is lower under restricted information in all three environments, with the largest decline in debt rollover (Table 3). This pattern supports the role of visible commitments in sustaining participant-led coordination.

Accordingly, we replaced “Communication trumps compensation” with a more specific discussion of sharing commitment information in Section 4.2. The revised explanation focuses on how commitments stabilize the funding position and how their visibility supports other participants’ decisions.

W2: Agent behavior and organization of Section 4

Response: We agree that the outcomes should be connected more clearly to agents’ decisions. We have expanded the analysis using two complementary sources of evidence: financial trajectories and recorded decision explanations.

Decisions and financial outcomes across rounds. In Section 4.2, we track when agents commit, how much funding is committed, how much remains free to exit, and how subsequent withdrawal, rollover, or pledge decisions change. We connect these decisions to the institution's cash and assets in bank runs, the early repayment deadlines in debt rollover, and the funding goal in crowdfunding (Figure 4).

The updated Figure 7 includes 140 episodes per environment. Comparing maximum committed shares below 20% with shares of at least 80% during rounds 1–3, bank-run success is 35.48% versus 88.57%, and debt-rollover success is 14.89% versus 80.65%.

What agents say about their decisions. Appendix I examines the explanations returned with agents' decisions. We compare explanations accompanying low and high withdrawal probabilities and examine references to locked funds, commitments, and previous redemptions. This complements the recorded action sequence with the considerations that agents explicitly mention.

We also reorganized Section 4 to separate the presentation of results from their interpretation:

  • Section 4.1: aggregate outcomes and the direct-action check.
  • Section 4.2: behavioral interpretation and the information comparison.
  • Section 4.3: robustness and ablation experiments.

This reduces repeated findings and makes the connection between results and explanations easier to follow.

AuthorsOpenReview ↗

Response to reviewer qbAF (Part3/4)

Response to Questions

Q1: Have you tested whether the baseline failure rates and mechanism rankings hold under prompt variations?

A1. Yes. We tested two kinds of prompt changes. Meaning-preserving rephrasing rewrites the original role, financial-rule, and mechanism descriptions using different wording while keeping their meaning, numerical values, and available information. Optimistic and pessimistic nudges add text suggesting that economic conditions are favorable or unfavorable, without changing the actual financial parameters. We ran 540 additional episodes across GLM, Qwen, and DeepSeek, all three environments, and M0–M3, using one rephrased version and the two nudge versions. Mechanism rules remain fixed throughout.

Meaning-preserving rephrasing. Each of M1–M3 still achieves higher mean welfare than M0 in all three environments. M3 retains the highest mean welfare in bank runs and crowdfunding. Debt-rollover rankings are more sensitive: within this three-model comparison, M3 welfare falls from 75.81% to 23.68% on the normalized percentage scale, while M2 becomes the best-performing mechanism. Thus, the improvement over M0 persists, although the strongest mechanism can change.

Optimistic and pessimistic nudges. We append an economic description to the original prompt. For bank runs, it describes the economic outlook as “favorable” or “unfavorable,” together with stronger or weaker confidence in financial institutions. The descriptions vary across the three scenarios, but returns, interest rates, cash, and other financial parameters remain unchanged. This checks how verbal suggestions about the environment affect agents' choices. The table reports M0 failure rates:

EnvironmentOriginal promptRephrasedOptimisticPessimistic
Bank run60.00%80.00%60.00%100.00%
Debt rollover93.33%100.00%66.67%100.00%
Reward crowdfunding13.33%13.33%0.00%100.00%

Baseline failures remain frequent in bank runs and debt rollover, while crowdfunding responds more strongly to the optimistic or pessimistic descriptions. Across all four prompt versions, M3 retains the highest mean welfare in bank runs and crowdfunding; M2 improves on M0 in debt rollover and leads under the three revised versions. Each cell summarizes 15 episodes, five per model. Prompt examples and the result table appear in Section 4.3 and Appendix J.5 (Tables 17–18).

Reviewer_NcM4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FRAIL evaluates LLM agents in dynamic bank-run, debt-rollover, and reward-crowdfunding environments. It compares compensated commitments, centralized agreements, and participant-led coalitions with a no-mechanism baseline. Across seven models, experiments find frequent collective failures and improved welfare under commitment mechanisms. Additional analyses examine early participation and robustness to changes in financial parameters.

Critical strengths
  1. The framework captures meaningful financial feedback: withdrawals trigger discounted asset sales, while promised compensation increases liabilities. This provides a useful setting for studying how individual actions and institutional rules jointly affect system stability.
  2. The study spans three distinct funding structures and seven models. The rise in bank-run survival from 23% to 74% under coalitions, together with the economic-regime checks, supports the effectiveness of the tested commitment mechanisms.
Critical weaknesses

The crowdfunding M1 budget is inconsistent across the specification. Table 4 and the prompts give the founder a margin of 0.05P, but Table 4 sets over-promise failure at bonuses exceeding 0.95P. The implemented rule and its effect on the reported M1 comparison are unclear.

Questions for authors

Was the crowdfunding M1 compensation limit 0.05P or 0.95P? Please clarify the founder payoff and failure rule used in the experiments, and whether resolving this discrepancy changes the reported M1 conclusions.

2 comments
AuthorsOpenReview ↗

Response to reviewer NcM4

Thank you for recognizing the financial feedback captured by FRAIL and for pointing out the inconsistency in Table 4. We checked the implementation and the recorded runs. The implemented M1 compensation budget was 0.05P0.05P, and the reported M1 results and conclusions remain unchanged. Here, PP is the pool at launch; the 0.95P0.95P entry in the original table was incorrect.

The experiments used a project cost share of κ=0.95\kappa=0.95, leaving a gross margin of (1−κ)P=0.05P(1-\kappa)P=0.05P. If BB denotes the total promised bonuses, the founder receives 0.05P−B0.05P-B upon successful reward delivery, before subsequent interest accrues.

We checked the script crowdfunding/sim/mechanisms.py (you can find it in the supplementary material "FinFrail.zip"), which computes the available margin and tests whether the promised bonuses exceed it:

def margin_of(cfg, P: float) -> float:
    return (1.0 - cfg.kappa) * float(P)

def is_overpromise(state: GameState, cfg, P: float) -> bool:
    return bonus_liability(state) > margin_of(cfg, P) + cfg.eps

This check is applied at both launch and reward delivery. Exceeding the budget triggers campaign failure; cfg.eps provides numerical tolerance.

We also checked the saved configurations and payment records of all 35 main-experiment M1 crowdfunding episodes—seven models with five episodes each:

What we checkedWhat we found
Compensation budget used in the experimentsAll 35 episodes used κ=0.95\kappa=0.95, giving a compensation budget of 0.05P0.05P.
Payment actually received by the founderIn all 34 successfully settled episodes, the founder received 0.05P−B0.05P-B at reward delivery, as specified above.
Final outcomes34 episodes launched successfully; the remaining episode did not launch. No episode failed because bonuses exceeded the budget.

The reported M1 outcome therefore remains 34/35 successful launches, with mean normalized welfare of 0.819.

We have corrected the budget entry in the original Table 4, now Table 5 in Appendix D.1, and added an explicit explanation of the founder’s gross margin and net payment immediately below the table. The 35-episode budget and payment check (FRAIL_rebuttal_evidence/checks/crowdfunding_budget_35.csv) lists the launch pool, compensation budget, promised and paid bonuses, and founder payment for each episode. The check guide (FRAIL_rebuttal_evidence/checks/README.md) explains the columns and links to the recorded event extracts and verification code.

We hope this clarifies the discrepancy and would be happy to answer any remaining questions.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer NcM4,

Thank you again for flagging the crowdfunding budget discrepancy. We have addressed your question and shared the results of our checks. We would be happy to discuss any remaining questions.

Best regards,
The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions

Thank you for recognizing the importance of studying collective financial failures among LLM agents and for the constructive feedback. We have strengthened the evidence and clarified the interpretation in response in the revised manuscript. The revised parts are marked in blue in the pdf. The main revisions are:

  1. Direct action selection (Section 4.1; Appendix J.3). We changed the prompts so that models choose a financial action directly, such as withdrawing or keeping a deposit, instead of reporting a probability for the simulator to sample. We provide the execution code and recorded multi-round examples.
  2. Mechanism controls (Sections 4.2–4.3; Table 3; Appendix J.4). We replaced the central LLM's decisions with fixed rules and separately removed extra information about other participants' commitments, while retaining the commitment and financial rules.
  3. Prompt and initialization checks (Section 4.3; Appendices J.5–J.7). We reworded instructions without changing their meaning, added optimistic or pessimistic descriptions without changing financial parameters, replaced simulation-related wording, and changed how initial deposits are distributed among depositors.
  4. Behavioral analysis and presentation (Section 4.2; Figures 4 and 7; Appendix I). We examined when agents commit funds, how later financial decisions change, and what agents say about those decisions, and reorganized the Results section.

We also corrected the crowdfunding budget description in Appendix D.1 and clarified why we ask agents to report action probabilities. The "reading_index.md" in the additional evidence package ("FRAIL_rebuttal_evidence.zip") maps the responses to code, trajectories, and the two numerical checks. The additional evidence package contains the corresponding files and can be downloaded directly from the supplementary material link.

AuthorsOpenReview ↗

Announcement to Reviewers and Area Chair

Dear Reviewers and Area Chair,

Thank you for your thoughtful feedback and for the opportunity to strengthen the paper. FRAIL studies how individually protective decisions can create collective financial fragility among LLM agents, and how commitments can help stabilize their interactions. The discussion helped us refine the interpretations of the results and strengthen our statement with clearer comparison. The revisions we made are listed below:

We expanded the analysis of how failures and stabilization unfold. The revised Section 4 added analysis on how mechanisms change the reasoning of agents, and eventually their actions. In bank runs, 35% of episodes with early committed shares below 20% survived, compared with 89% of episodes where the share reached at least 80% in the first three rounds. We also used text analysis to show that activated commitments are systematically taken into consideration when agents are making supportive moves (choose to not withdraw/ rollover the debt/ pledge the funding).

We added comparisons that clarify how commitment mechanisms work. We replaced the central LLM with fixed rules for organizing offers and agreements and found that the enhancement still exists. We also masked additional information (other participants' commitments) to examine how much the stabilizing signal matters to the outcome. [Section 4.2]

We tested the main findings across different ways of eliciting decisions and presenting financial information. Collective failures and the gains from commitment mechanisms remain evident when agents choose their actions directly. We tested variations of the experiments including prompt rephrasing, optimistic and pessimistic economic descriptions, neutral financial wording, different initial deposit allocations, and more favorable or tighter financial conditions. Also we tested the robustness of our results by giving explicit guidance to the player, in order to show the fragility is not purely driven by misreading of the agents.

These additions make the revised paper a more complete account of how collective fragility develops and when commitments can help. We appreciate the reviewers' questions, which motivated us to examine the agents' behavior more closely and made the contribution of the paper clearer.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.