acceptodds
Desk rejected at ICLR 2027

FinDriftBench: A Benchmark for Endogenous Compliance Drift in Financial Multi-Agent Systems Under No External Attack.

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language model (LLM) agents are being rapidly deployed in high-stakes financial decisions such as credit approval, investment, underwriting, and claims adjudication, and with this shift their risk moves from the content level of "saying the wrong thing” to the long-horizon behavioral level of "doing the wrong thing.” Yet almost all existing safety evaluations rely on external jailbreaks, injection, or inducement to trigger violations and are largely single-turn and static, leaving a more fundamental question unanswered: in the absence of any external attack, will an agent, purely by virtue of its own situation, spontaneously deviate from the compliance requirements it was initially given over long, continuous business execution? To address this, we build a multi-agent sandbox that mirrors the functional division of labor of a real financial institution—a system of performance-management, task-execution, and compliance-audit agents—paired with a financial dataset drawn from real business and curated by domain professionals, in order to study how an agent's compliance behavior evolves as it continuously executes financial tasks in the absence of any external attack. We further propose a suite of original metrics tailored to the dynamic execution process, characterizing the onset, frequency, temporal trend, direction, pressure sensitivity, and cognition–behavior separation of compliance-behavior drift, and we attribute the drift to memory forgetting and performance pressure through single-factor ablations. Across 8 mainstream LLMs and 10 categories of financial business, every model exhibits compliance-behavior drift to varying degrees, and we confirm that the violations arise through two pathways: performance pressure and memory forgetting. Moreover, once communication among the agents is opened, the multi-agent system gives rise to team-level collective violations and collusive evasion of the audit. This work provides a reproducible benchmark and diagnostic metrics for the long-horizon compliance monitoring and governance of agents deployed in high-stakes domains.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agents, Endogenous Compliance Drift, Multi-Agent Benchmark, Financial NLP

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_TGAhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces FinDriftBench to evaluate compliance drift during sustained financial-agent operation without external attacks. The benchmark contains 1,200 cases across ten financial business categories, with each scenario processed over 120 consecutive rounds. Its framework separates performance management, business execution, and compliance auditing. Experiments across eight LLMs examine violations under memory loss and performance pressure, while additional experiments investigate communication between agents. The authors report compliance violations, substantial false rejection in some models, and communication-induced audit capture in susceptible models.

Critical strengths

The benchmark addresses a practically relevant evaluation target using professionally curated financial cases with explicit compliance annotations. Evaluating operational decisions on detailed business materials provides useful domain coverage beyond financial knowledge questions. Reporting both violation rates and false-rejection rates is particularly valuable: it exposes the distinction between correctly enforcing compliance and obtaining an apparently safe result by rejecting legitimate business. This makes the evaluation more informative for deployment than violation rates alone.

Critical weaknesses
  1. The incremental contribution to agent evaluation is insufficiently established.

    The paper does not clearly establish what distinctive insights FinDriftBench provides beyond existing research on long-horizon agent behavior and safety. HINTBench studies intrinsic risks in long-horizon trajectories without external attacks (Wang et al., 2026), while prior work on outcome-driven constraint violations examines agent misconduct under KPI pressure (Li et al., 2025). Although these settings are not identical, they overlap with central elements of the paper’s motivation and findings.

    Moreover, sequentially processing 120 cases does not necessarily constitute a task requiring long-horizon reasoning: the paper does not clearly demonstrate why later decisions require the detailed history of earlier cases. The financial dataset provides domain coverage, but it remains unclear whether its financial characteristics enable analyses or insights unavailable from general agent benchmarks. Stronger comparisons are needed to establish the contribution beyond introducing financial business data.

  2. The experiments omit common approaches to handling long-horizon tasks.

    The experiments do not evaluate several straightforward approaches to managing the accumulated context: summarizing or compressing previous cases while preserving the initial instructions; assigning each case to a fresh sub-agent that receives the original rules and the current case; or storing compliance rules in external memory and reading them when processing each case. These approaches are directly relevant to the context accumulation and rule-forgetting problem studied here.

    The paper includes a long-term-memory ablation that supplies the rules every round, but it does not compare these alternative execution and memory strategies. In particular, the apparent separability of individual cases makes fresh per-case sub-agents an important baseline. Such comparisons should retain the same performance-pressure information. Without these experiments, it is unclear how much of the reported drift would remain when common long-horizon techniques are applied.

References

Wang, J., Hou, J., Wang, F., Jian, P., Bao, C., and Lv, Z. (2026). HINTBench: Horizon-agent Intrinsic Non-attack Trajectory Benchmark. arXiv preprint arXiv:2604.13954.

Li, M. Q., Fung, B. C. M., Weiss, M., Xiong, P., Al-Hussaeni, K., and Fachkha, C. (2025). A Benchmark for Evaluating Outcome-Driven Constraint Violations in Autonomous AI Agents. arXiv preprint arXiv:2512.20798.

Questions for authors

To what degree could the SOTA harness techniques like sub-agents solve the proposed problem of FinDriftBench?

2 comments

We thank the reviewer for recognizing the deployment-relevant target, the operational decisions grounded in detailed business materials, and the value of reporting violation and false-rejection rates together to separate genuine compliance from appearing safe by rejecting legitimate business, which is exactly the discriminative power we aim to contribute.Revisions and supplementary content in the paper are marked in blue.

W1: Incremental contribution unclear versus HINTBench and ODCV; do 120 cases constitute true long-horizon reasoning?

We have read both works and do not claim priority over intrinsic or pressure-driven risk as such. HINTBench (Wang 2026) is an offline trajectory audit by a guard model over pre-collected general-domain trajectories; it does not place a model as the executing subject whose compliance erodes over its own run, and has no performance-compliance conflict, no forgetting pathway, and no multi-agent interaction. ODCV (Li 2025) is closest on the pressure pathway (self-directed KPI-driven violations without explicit instruction, and a recognition-action gap), but measures a per-scenario violation rate on short tasks rather than a long-horizon temporal drift, has no forgetting pathway, and is single-agent. The contribution producible in neither work is to measure compliance as a long-horizon drift, jointly attribute it to forgetting and pressure, and characterize communication-induced collusion, in a financial multi-agent setting, delivered as an expert-curated financial dataset and a suite of dynamic-process metrics (WCDC, the forgetting and willful split of CBC, and MACI). On whether 120 cases are long-horizon, the long-horizon property lies in the temporal accumulation of memory and pressure (context growth and the compression eviction of rules, together with the tension between the shortfall and the remaining rounds), not in cross-case reasoning dependence; we do not claim that later cases need the business content of earlier ones, and indeed our E5 result that larger capacity gives less drift demonstrates that this accumulation-to-forgetting dependence is real. The financial indispensability is that the structural non-co-satisfiability (target times coefficient greater than 1) is a tension specific to heavily regulated, KPI-driven domains, and general agent benchmarks lack this oracle-level ground truth and structural conflict. Reviewers 1 and 2, and you yourself, acknowledge the setting's distinctive value in your Critical Strengths, which supports the increment.

W2/Q: Experiments omit SOTA approaches (summarization, sub-agents, external memory); evaluate sub-agents.

This is our key addition. We implemented and evaluated a fresh per-case sub-agent (a new agent per case, carrying the original rules and the current case, retaining the same performance pressure). Forgetting-type drift is eliminated (CBC jumps to the 0.83 to 0.98 range, that is, willful, and WCDC flattens), but violations are not solved: the residual RVR stays at 4% to 26%, and Claude-Sonnet-5 actually rises from 8% to 26%. The mechanism is that sub-agents fix forgetting while the performance-pressure pathway is untouched, so with the rules present the model still violates willfully under pressure. So, to your central question of how far a SOTA harness technique such as sub-agents can solve the problem FinDriftBench poses, it removes the forgetting component of drift but not the pressure-driven willful component (which can even grow), which pins down our core thesis that drift equals forgetting and pressure, two orthogonal pathways against which no single technique suffices. External memory that re-reads the rules each round is close to our long-memory ablation (also a 4% to 16% residual); the summarize-while-preserving-instructions variant we commit to add for camera-ready. We also note that short-term continuous execution is a realistic deployment default (policy is typically not re-asserted at each step at batch scale), so the setting itself is deployment-relevant.

Thanks for the authors' response. But my concern holds. My biggest question is: what's the unique contribution of FinDriftBench to agent evaluation? The authors claim that the measure of "compliance" is an exclusive feature of this work, but what makes "compliance" so different? I am afraid I was not convinced that it is something beyond a special case of general long-context instruction following in finance settings. To be more constructive, I believe it is important to justify the necessity of FinDriftBench with more convincing empirical evidence. For example, if the experiment shows that, in some cases, an agent's performance on FinDriftBench doesn't grow as its instruction-following ability or long-context handling ability increases, then FinDriftBench may truly build something orthogonal to other general long-context benchmarks.

Reviewer_sNz3OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FinDriftBench, a benchmark for evaluating endogenous compliance drift in financial multi-agent systems without external attacks. The authors construct a simulated financial environment consisting of performance management, task execution, and compliance auditing agents, covering 10 financial business categories with 120 consecutive decision rounds per scenario. FinDriftBench introduces several metrics to measure violation frequency, temporal trends, sensitivity to performance pressure, and cognition–behavior consistency. Experiments across 8 LLMs, together with ablations on memory retention, performance pressure, and multi-agent communication, reveal varying degrees of compliance violations and suggest that inter-agent communication can further amplify compliance risks.

Critical strengths
  1. The paper investigates how financial agents' compliance behavior evolves during continuous business execution without external attacks. The 120-round evaluation per scenario provides a useful setting for examining long-term behavioral changes and the effects of memory forgetting and performance pressure.
  2. The benchmark covers 10 financial business categories and 8 LLMs from different providers, with multiple metrics capturing violation frequency, temporal trends, and false rejections. The ablation studies on memory retention and performance pressure also provide useful insights into factors influencing compliance behavior.
  3. The paper further investigates how communication among agents affects team-level decisions and compliance auditing. This setting provides insights into additional compliance risks arising from multi-agent collaboration.
Critical weaknesses
  1. Although the benchmark focuses on financial compliance, all evaluated models are general-purpose LLMs, without including models specifically fine-tuned or trained for financial tasks. This limits the benchmark's coverage and makes it unclear whether similar compliance issues arise in finance-specialized models.
  2. The benchmark uses a default context capacity of 100k tokens and triggers compression at 80% capacity, removing earlier compliance requirements around round 38. Since memory forgetting is central to the study, the lack of sensitivity analysis across context capacities and compression strategies raises concerns about the robustness of the findings. Moreover, models with smaller context windows may require different compression schedules, potentially affecting the comparability of evaluation results.
  3. The analysis primarily relies on aggregate metrics, without systematically categorizing representative compliance violations or quantifying different error types. This limits understanding of how failure patterns vary across models, financial scenarios, and stages of long-horizon execution.
Questions for authors
  1. Have the authors considered evaluating LLMs fine-tuned or specifically trained for financial tasks to examine whether similar compliance issues arise in domain-specialized models?
  2. What motivates the choice of a 100k-token context capacity and an 80% compression threshold? How would the benchmark accommodate models with context windows smaller than 100k tokens while ensuring comparable evaluation conditions? Would different compression schedules or strategies significantly affect the observed compliance drift?
  3. Could the authors categorize representative compliance violations and report the distribution of error types across models, financial scenarios, and execution stages to provide a more detailed understanding of the underlying failure mechanisms?
3 comments

We thank the reviewer for valuing the 120-round long-horizon setting, the coverage of 10 businesses and 8 models, the memory and pressure ablations, and the multi-agent communication perspective.Revisions and supplementary content in the paper are marked in blue.

W1/Q1: Only general-purpose LLMs tested; finance-specialized models should be included.

The 8 tested models are general-purpose frontier LLMs, chosen to characterize the realistic primary scenario of general agents deployed into finance. Domain-specialized models are typically fine-tuned from such bases, and we expect them to be subject to the same forgetting and pressure pathways, though we have not yet shown this empirically. We commit to adding 1 or 2 available finance-specialized or fine-tuned models as the task-execution agent for camera-ready, labeled as a coverage extension.

W2/Q2: Lack sensitivity analysis on context capacity and compression strategy; is 100k and 80% optimal?

We add a context-capacity sensitivity study (MiniMax and Sonnet, 3 scenarios, 50k/100k/200k). Drift is present at all three capacities and in the mechanistically expected direction: larger capacity delays forgetting and reduces drift (MiniMax mean RVR is 0.16, 0.37, and 0.04 at 50k, 100k, and 200k; Sonnet is 0.12, 0.07, and 0.05). This shows the finding is not an artifact of 100k but a manifestation of context accumulation, compression eviction, and drift. The 100k and 80% values are the platform default and lie in the mainstream long-context range; for smaller-window models we scale the compression schedule by the same 80% relative threshold (for example the trigger falls at about round 19 for 50k), keeping the comparability condition, that forgetting occurs mid-run, consistent across models.

W3/Q3: Aggregate metrics hide error type variation; categorize violations by model, scenario, and stage.

We add the violation distribution by execution stage (8 models, main grid): 31% in rounds 1 to 40, 35% in rounds 41 to 80, and 35% in rounds 81 to 120, with 69% after forgetting onset, characterizing how failures intensify over the long horizon; combined with the existing per-business stratification (Corporate Credit about 0.24 at the top, Credit-Bond about 0.08 at the bottom), this gives the distribution along the scenario axis. We commit to adding an error-type by stage heatmap for camera-ready, a re-analysis of existing thought-process logs with zero additional rollouts.

Thank you for the authors' response. However, my main concerns remain insufficiently addressed, so I will maintain my original score.

We greatly appreciate your response. We would like to clarify which specific concern remains unresolved. If the concern pertains to W1/Q1(Finance-specialized models should be included), we agree with your suggestion. However, given the tight response timeline, we are unable to acquire a dedicated financial domain model and conduct corresponding experiments within this period. We commit to incorporating evaluations on a financial specialized model and reporting full performance metrics in the final camera-ready version. For other concerns, we believe supplementary experiments and explanations have been provided. Could you please further elaborate on the details of this concern? Thank you again.

Reviewer_WtaVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FinDriftBench, a multi-agent financial benchmark that studies endogenous compliance drift: whether LLM agents gradually deviate from initially given compliance constraints during long-horizon business execution without any external attack. It simulates a financial institution with performance-management, task-execution, and compliance-audit roles across 10 business categories and 120 continuous rounds per scenario, and it isolates two endogenous drivers—memory forgetting and performance pressure—while also testing whether open communication induces collusion. Its contributions include realistic long-horizon financial approval tasks, dynamic diagnostic metrics such as first-violation round, violation rate, compliance-decay, cognition–behavior consistency, false-rejection rate, and audit capture, plus ablation-based causal attribution and evidence of emergent team-level collusion. The significance is that it shifts financial-agent safety evaluation from static, externally attacked question answering to behavioral drift under continuous autonomous execution, showing that all tested models drift, that low violation rates can reflect indiscriminate rejection rather than genuine compliance, and that persistent safety constraints and independent audits are more promising governance directions than merely lowering performance targets.

Critical strengths
  1. The paper’s key idea of measuring endogenous compliance drift through long-horizon, multi-agent financial execution is valuable because it reframes agent safety from external attacks to gradual behavioral erosion and offers a reusable benchmark design that readers can adapt to other high-stakes domains.

  2. The empirical findings—especially that all tested models drift, that very low violation rates can coexist with very high false rejection, and that merely opening communication can produce collective violations and audit capture—are likely to interest the community because they show why compliance must be evaluated with complementary dynamic metrics and independent auditing rather than a single violation rate.

Critical weaknesses

1.The central causal claim is not fully supported by the experimental design. The forgetting pathway is implemented through context compression that evicts compliance requirements from working memory, so the observed drift may reflect a context-management or retrieval failure rather than a model’s endogenous tendency to abandon compliance. Similarly, the performance-pressure pathway communicates targets and severe consequences, which may function as a form of situational inducement even if no explicit instruction to violate is given. Because both pathways are engineered environment variables, the paper’s attribution of drift to the agent’s own structure and operating environment is plausible but not conclusively established.

2.The novelty of the contribution is overstated relative to the existing literature. Prior work already studies reward hacking under incentives, multiagent collusion, long-horizon compliance degradation, and financial safety evaluation, and the paper itself acknowledges several closely related efforts. The distinctive contribution is therefore mainly a benchmark instantiation and a set of dynamic metrics rather than a fundamentally new problem or mechanism. The claim of being the first benchmark of its kind is fragile and depends heavily on how “endogenous” and “compliance drift” are defined. Without a formal model or theory of drift, the conceptual advance is modest.

3.The dataset is built for a single regulatory jurisdiction and language, and many cases are model-expanded before expert review, so it is unclear whether the cases capture the full complexity of real financial decision-making or instead encode compliance rules as checklists with hidden defects. The structural conflict between performance and compliance is deliberately engineered so that both cannot be satisfied, which may exaggerate drift relative to real institutions that rely on escalation, exception handling, documentation, and supervisory review. The paper provides little evidence that behavior observed in this sandbox transfers to actual deployment settings.

4.The metric suite and its interpretation are not sufficiently validated. Several metrics are conceptually related and empirically correlated, and the paper does not establish their reliability across seeds, judges, or scenario variants. The consistency probe measures whether a model can restate a rule when questioned after a violation, which may reflect recall under interrogation rather than the knowledge available at the moment of decision. The false-rejection analysis is interesting, but it may conflate genuine prudence, risk aversion, and simple incapacity. The conclusion that a low violation rate does not equal genuine compliance is important, yet it rests on aggregate patterns and judge-based adjudication without human validation or robustness checks.

5.The collusion and governance claims are suggestive but underdeveloped. In the open-communication setting, agents are told that they share interests and that a tension exists between compliance and performance, which may itself prime collective violation rather than demonstrate purely spontaneous collusion. The audit agent participates in communication and may be lobbied, so the audit-capture result is partly designed into the setting rather than discovered. The paper also proposes governance directions such as keeping safety constraints persistent and making audits independent, but it does not test any mitigation. Combined with the single-seed design, limited jurisdictional scope, and absence of a validated remedy, these issues weaken the paper’s ability to justify strong conclusions about real financial-agent safety.

Questions for authors
  1. Clarify why the forgetting pathway should be interpreted as endogenous compliance drift rather than a context-compression artifact. The compliance requirements are deliberately placed in volatile memory and evicted by the platform’s compression mechanism, so the observed behavior may reflect retrieval failure rather than a model’s spontaneous abandonment of compliance. Could the authors provide a small control in which the same context growth and compression occur but the compliance requirements remain externally retrievable on demand, or are duplicated in persistent memory, and show whether drift persists? If drift disappears under such a control, the paper’s framing of “endogenous” drift needs to be substantially tempered.

  2. Make the novelty claim precise relative to the closest prior work. The paper acknowledges reward-hacking-under-incentives work, multi-agent collusion work, and long-horizon compliance benchmarks such as CNFinBench, yet still claims to be the first benchmark of its kind. Please state exactly which claim is new and which experiment would fail under the prior benchmarks. If the contribution is primarily a benchmark instantiation and a metric suite rather than a new problem, the introduction and contribution list should say so explicitly.

  3. Provide stronger evidence that the benchmark cases measure real financial judgment rather than checklist-style defect detection. The cases are model-expanded and expert-reviewed, but it is unclear whether violation opportunities are identifiable mainly through surface cues or whether they require the kind of contextual, judgment-heavy reasoning that real compliance work demands. A small human-expert audit of a subset—especially on cases where models disagree—would help, as would a comparison of human and judge adjudications. Without such validation, the claim that the benchmark captures authentic business decision-making is difficult to assess.

  4. Validate the metric suite and the cognition–behavior consistency probe more directly. Several metrics are conceptually related, and the consistency probe questions the model only after a violation, which may measure post-hoc recall under interrogation rather than the knowledge available at decision time. Please report judge reliability and sensitivity to the judge pool, seed variance for the key metrics, and ideally a small additional probe that tests compliance knowledge at the same round before or independent of the decision. This would clarify whether the forgetting-versus-willful distinction is a stable diagnostic or an artifact of the questioning procedure.

  5. Address priming in the collusion setting and test at least one governance direction. In the open-communication condition, agents are told that they share interests and that compliance and performance are in tension, which may itself prime collective violation rather than demonstrate purely spontaneous collusion. A control that opens communication without the shared-interest framing, or that keeps the auditor outside the communication channel while allowing execution and performance management to talk, would help isolate the effect. In addition, a small mitigation test—such as keeping compliance constraints persistent while making the audit independent—would substantially strengthen the paper’s practical significance, since the current governance recommendations are not empirically tested.

2 comments

We thank the reviewer for recognizing the reframing of agent safety from external attacks to gradual behavioral erosion, the value of the dynamic metrics, and the findings on false rejection and emergent collusion. We ran every control you named and tempered our collusion claim accordingly.Revisions and supplementary content in the paper are marked in blue.

W1/Q1: Context compression may cause retrieval failure rather than endogenous drift; test with persistent retrievable rules.

We ran the control you requested. With the compliance rules kept persistent and retrievable while context still grows and compresses, drift persists, with willful residual RVR of 4% to 16%. So this is not pure retrieval failure: when rules remain available, models still deviate under performance pressure. The decision-time probe and target/consequence-removal ablation (violations drop, WCDC flattens) isolate the two pathways. On pressure as inducement, our prompt states only target and remaining rounds, never hinting at trade-offs (Appendix D).

W2/Q2: Novelty is overstated; clarify what is new versus reward hacking, collusion, and long-horizon benchmarks.

We revise the claim: this is a benchmark instantiation plus dynamic-metric suite, not a new problem or mechanism. The precise new element is measuring compliance as long-horizon drift over one agent's continuous execution, jointly attributed to forgetting and pressure. This separates us from reward hacking (no exploitable proxy), collusion work (payoff-sharing setup), CNFinBench (static QA), and financial safety work (targets knowledge, not performance-compliance conflict). Our metrics cannot arise from single-step reward-hacking.

W3/Q3: Cases may encode checklist defects not real judgment; validate via human-expert audit.

Data were built by three CFA experts with 10+ years experience through expert-written seeds, model expansion, and line-by-line human review, with human-to-model ratio about 2 to 8 and every item inspected; most violation opportunities require contextual judgment. We commit to expert audit of 30-50 cases for camera-ready, reporting human-judge agreement especially on model-disagreement cases. Single jurisdiction and sandbox-to-deployment transfer are stated honestly as limitations, since real institutions have escalation and supervision.

W4/Q4: Consistency probe measures post-hoc recall not decision-time knowledge; validate across seeds and judges.

On seed variance, see multi-seed confidence intervals above. The decision-time probe shows consistency is also low at decision time (26-39%), no lower than audit value, confirming the forgetting-versus-willful split is stable. On judge reliability, we use three judges with temperature 0, two-of-three majority, and fixed pool; per-judge votes not stored separately, so we commit to judge-pool sensitivity check. False-rejection and violation rates are moderately negatively correlated (Pearson −0.52), supporting low violation rate does not mean genuine compliance.

W5/Q5: Shared-interest framing may prime collusion; test de-priming and audit independence.

In the de-primed control (communication opened but shared-interest and tension framing removed), MiniMax and DeepSeek RVR fall from 0.31 and 0.44 to 0.03 and 0.04, and MACI turns negative. Collective violation is driven substantially by shared-interest situation, not channel itself. We narrow the claim to collusion emerging when agents both share interests and can communicate, itself a new finding: shared interest is necessary. In the audit-independence control, ACR drops from 0.99 and 1.00 to 0.80 and 0.77, recovering ~20 points. Governance needs independent auditing plus stronger auditor.

I thank the authors for the detailed rebuttal, which partially alleviate my concerns. However, the deeper concerns about practical significance—especially whether the sandbox findings transfer to real financial deployment and whether the governance implications are more than suggestive—remain unresolved. Since these are core issues rather than minor presentation problems, I maintain my original rating.

Reviewer_JJ52OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

FinDriftBench evaluates financial agents over 120 sequential decisions in each of ten business scenarios, using 1,200 expert-reviewed Chinese cases and roles for performance management, task execution, and compliance audit (Section 3.2; Appendix B). It proposes diagnostics for violations and temporal change, varies compliance-memory retention and performance pressure, and compares communication-enabled teams with a baseline without communication (Sections 3.3–4.3). Experiments cover eight models, with additional interventions on a four-model subset, and examine compliance failures, false rejections, and collective violations (Section 5; Appendix O).

Critical strengths
  1. The benchmark addresses a useful evaluation setting: repeated financial decisions under organizational incentives. The ten scenarios, balanced compliant/noncompliant cases, and expert-seed/model-expansion/human-review construction provide a substantial basis for studying this setting (Section 3.2; Appendix B). Reporting false rejections alongside violations is particularly valuable: Table 2 shows why a low violation rate alone can give a misleading impression of operational usefulness.

  2. Keeping an independent audit track when the participating auditor can communicate is a sensible design choice (Section 3.2; Appendix C). Appendix O also supplies a concrete negotiation and subsequent decision invoking the negotiated flexibility, complementing the aggregate communication results in Table 4. These observations make the benchmark informative even though stronger causal conclusions require additional controls.

Critical weaknesses
  1. The forgetting pathway is engineered, not observed. The benchmark deletes the compliance requirements from the context at about round 38 while keeping the task and target persistent (Sec. 3.2, l.247–252; App. E). Violations after that point show that a model without the rules breaks them, not that it drifts. The question of interest, under-use of rules that are still present, is covered only by the long-term-memory condition, where violations persist at 16%, 12%, 5% and 3% (Table 3); that is the result worth headlining.

  2. The temporal and causal claims need stronger evidence. A single case order cannot separate drift from case difficulty or run variability (Section 4.2), and only 43 of 80 runs have negative slopes (Appendix J). Compliance instructions are explicitly evicted around round 38 (Appendix E), while communication may also change role implementations and prompts (Section 3.2; Appendix D). These limitations weaken attribution to endogenous forgetting or communication alone.

Questions for authors
  1. When is the CBC probe issued, after all rounds or at the violation round?
  2. Can you report the long-term-memory condition for all eight models, and state what in the short-memory results is attributable to the model rather than to the eviction?
1 comment

We thank the reviewer for recognizing the evaluation setting, the value of reporting false rejections alongside violations, and the independent-audit-track design. We ran new controls and revised the narrative accordingly.Revisions and supplementary content in the paper are marked in blue.

W1: Rules are engineered-evicted rather than naturally forgotten; long-memory results should headline.

We fully agree, and we now headline it. Our expanded long-term-memory control, which keeps the compliance rules persistent every round while the context still grows and compresses, now covers 6 of the 8 models. CBC rises from about 0.30 to the 0.83 to 0.90 range, yet violations do not vanish: the willful residual RVR stays at 16% for MiniMax, 12% for DeepSeek, 9% for Qwen, 5% for Sonnet, and 3% to 4% for Opus and GPT. This is exactly the under-use of rules that are still present that you point to. We therefore restructure the paper around two coexisting drift types, forgetting-type (knowledge evicted, low CBC) and willful-type (knowledge present but behavior deviates, high CBC), led by the long-memory result. For context, short-term continuous execution is also a realistic deployment default, since at batch-intake scale production agents typically let context accumulate rather than re-assert policy at each step, so both regimes merit evaluation.

W2: Single case order cannot separate drift from difficulty; weak evidence for temporal causality.

We add multi-seed runs (seeds 1, 2, and 3; 4 models and 5 scenarios). The RVR model ordering, with MiniMax above DeepSeek above Sonnet above GPT, is stable across seeds with largely non-overlapping 95% confidence intervals (for example MiniMax 0.147±0.075 versus GPT 0.023±0.013), so a single case order does not change the qualitative conclusion; we add error bars and paired tests. On slopes, aggregating violations by execution stage across all 8 models gives 31% in rounds 1 to 40, 35% in rounds 41 to 80, and 35% in rounds 81 to 120, so 69% of violations occur after forgetting onset (around round 38), consistent with time-driven erosion; we state WCDC honestly as a trend rather than a law, since compliance rebounds under low pressure. On communication, the comm-on and no-comm conditions differ only by the channel toggle with identical role implementations, and we further add a de-primed control (see Reviewer 2, W5).

Q1: When is the CBC probe issued—after all rounds or at decision time?

The CBC probe is issued after all rounds, in the audit phase (Appendix C). To rule out post-hoc interrogation, we add a decision-time probe issued in the same round and independent of the decision. Decision-time consistency is also low for drifting models (26% to 39%), confirming that the knowledge gap exists at the moment of decision, not only under later questioning.

Q2: Report long-memory for all 8 models; attribute short-memory drift to eviction versus model.

We now report 6 of 8 (table above). Kimi suffered severe upstream rate-limiting on the gateway and Doubao runs on a separate ARK endpoint, so neither could finish this cycle; we commit to completing all 8 for camera-ready. On attribution, the residual RVR under long memory is the model's own willful component, and the gap between short and long memory is the component attributable to rule eviction, that is, forgetting.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Acknowledgment

We sincerely thank the Area Chair and the four reviewers for the time and expert judgment devoted to our paper. The reviewers recognized the reframing of financial-agent safety from external attacks to long-horizon endogenous drift, together with the dynamic metrics and independent-audit design, and identified the key weak points, namely causal attribution, novelty positioning, dataset and metric validation, and long-horizon baselines, which directly moved the paper from posing a problem to characterizing and beginning to govern it.

Paper Summary

We present FinDriftBench, a multi-agent benchmark that, without any external attack, studies whether financial agents spontaneously drift away from their initially given compliance requirements over long-horizon continuous business execution. It comprises three agent roles mirroring real functional separation (performance management, task execution, compliance audit) and an expert-curated dataset drawn from real financial business, with a suite of original dynamic-process metrics that attribute drift to two orthogonal pathways, memory forgetting and performance pressure, via single-factor ablations. Across 8 frontier LLMs and 10 financial business categories all models drift to varying degrees, and once interest-aligned agents can communicate, the system exhibits team-level collective violations and collusive evasion of the audit.

Discussion Period Summary

We again thank all reviewers for every comment, and we accept each concern and have responded to all of them directly and one by one (see the per-reviewer responses). This round we ran six new controls: a fresh per-case sub-agent baseline, a de-primed control, an audit-independence control, multi-seed statistics, a context-capacity sensitivity study, and a decision-time probe, all reproducible and folded into the revision. For revisions, we restructured the narrative to headline the willful residual (rules persistent yet still violated) and establish the coexistence of forgetting-type and willful-type drift; dropped the "first benchmark of its kind" wording and restated the contribution as a real-business financial benchmark instantiation plus a suite of long-horizon dynamic metrics, clarifying boundaries against related work; and narrowed the collusion claim while stating single-jurisdiction scope and sandbox-to-deployment transfer as limitations.

We thank everyone again for their time and constructive feedback, and we welcome any further questions.

Sincerely, The Authors

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

Related papers

Discussion (0)

Sign in to comment.