acceptodds
Under review as a conference paper at ICLR 2027

AML-view: A Realistic Institution-Centric Synthetic Dataset and Generator for Anti-Money Laundering

Abstract

Anti-money-laundering (AML) investigation requires reasoning over a rich financial ecosystem using only one institution's evidence. Public data rarely expose this setting, while synthetic releases typically emphasize transactions or assume complete visibility. We present AML-view, a dataset and configurable generator that separates a shared, investigation-rich financial world from institution-local observations of it. AML-view integrates statistically grounded parties, ownership, accounts, due-diligence attributes, and payments, instantiates 38 documented laundering scenarios, and propagates illicit value through downstream transfers. Local views retain private records for customers but mainly transactional and public information for external counterparties. Each evaluated world contains about 10 million transactions over 70,000 accounts. AML-view covers all 13 audited investigation axes, versus 7 for the closest comparator, and passes all 35 population-calibration checks. By exposing matched institution-local views of the same financial world, AML-view reveals effects hidden by global-only evaluation: detection becomes harder for most models under realistic information boundaries, and the relative ranking of models can change. In blinded review, AML practitioners recovered laundering patterns from synthetic suspicious cases at a rate comparable to real cases, with recovery increasing from 0.500 using transactions alone to 0.765 with richer investigation context, while showing no consistent ability to distinguish synthetic from real cases. AML-view provides a large-scale, expert-validated, reproducible foundation for AML detection and investigation research.

Primary area:
datasets and benchmarks
Keywords:
synthetic data generation, anti-money laundering, graph dataset, partial observability, financial crime investigation, dataset

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_r9GzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

AML-view is a synthetic dataset generator to enable anti-money-laundering research. It creates people, companies, ownership links, bank accounts, customer records, and payments. It then produces a separate view for each bank, showing that bank’s transactions and limiting information about other banks’ customers. The authors check population statistics, compare the information available in different datasets, evaluate detection models, and report a study with four AML practitioners.

The paper offers a potentially useful research resource. However, its results mainly show that the generated population matches selected statistical targets and that some synthetic cases contain recognizable laundering patterns. They provide less evidence that the full transaction system behaves like a real financial system.

Critical strengths

The richer investigation data (compared to previous work) could support useful research. Linking customer records, households, ownership, and payments under explicit bank visibility rules could enable studies that transaction-only datasets cannot support. Releasing these elements together (along with the generator) is a meaningful potential contribution.

Critical weaknesses
  1. There is not enough evidence that the transactions and labels are realistic

The strong calibration results concern population and company statistics that were used to build the population. Matching them does not show that payment amounts, timing, counterparties, and ownership links interact realistically. The rules in Appendix C are plausible, but the evaluation does not sufficiently check the resulting financial behavior against independent references. Several ownership and laundering settings are design choices. Positive labels also depend strongly on a chosen rule: accounts spend illicit funds first, and a payment is positive if at least 25% of its value is illicit.

The practitioner study also does not validate these mechanisms. It tests selected case files, and the paper does not adequately explain case selection or how written answers were scored.

  1. Much of the modeling already exists, and the importance of the added features is not clear

The paper includes a financial world, tracking illicit funds, and separate bank views. AMLWorld and AMLGentex already describe substantial parts of this approach. The paper acknowledges some of this overlap but its statement that no existing release provides multiple bank views of one shared world is too broad. The clearest remaining contribution is richer investigation information and explicit rules for hiding private information about other banks’ customers. However, the work does not establish how important that extra coverage is.

Questions for authors
  1. What independent evidence shows that the generated payments and ownership relationships behave realistically, beyond matching population statistics?

  2. Are global and local models tested on exactly the same transactions and time intervals? How do their training data differ, and how are the out-of-window positive transactions handled?

  3. What important research task becomes possible with AML-view that the closest existing releases cannot support? Which result demonstrates this?

  4. How were practitioner cases selected and written answers scored? Were the scorers blinded to case origin and information condition, and was agreement between scorers assessed?

  5. Do Investigator’s strong detection results hold at the original laundering prevalence after removing possible shortcuts in the synthetic data, and what explains its success given the difficulty of real-world AML?

  6. How many generator parameters are set by the authors rather than estimated from data, how were they selected, and how sensitive are the detection results to those choices?

8 comments
AuthorsOpenReview ↗

Response to Reviewer r9Gz (1/8): W1, realism of transactions and labels

We thank the reviewer for the careful review and for recognizing the value of releasing linked investigation records with the generator.

We agree that the calibration checks measure agreement with the targets used to construct the population, not realism beyond them, and Section 4.2 states this. The remaining design choices are declared in Section 3.1 and Appendix F.5, and the labeling rule is specified in Appendix D.3. Ownership is only partly a design choice: the ownership thresholds follow legal sources and the transfer rates follow published statistics (Tables 31 and 32). The submission, however, did not check the resulting financial behavior against independent references, and it did not explain how practitioner cases were selected and scored. We therefore added four analyses.

Fidelity checks beyond calibration targets. Over five seeds (details in Q1), log transaction amounts are close to those of AMLWorld (Kolmogorov–Smirnov distance, KS, of 0.030, where 0 means identical distributions). As a descriptive check, the first-digit frequencies have a mean absolute deviation of 0.008 from Benford's law, below the 0.012 acceptability threshold. AMLWorld is itself synthetic, so this comparison measures agreement with another synthetic benchmark, not with real payments.

Ownership structure. AML-view ownership groups contain 2.35 legal units on average, compared with about 3.5 in French enterprise groups (Insee, 2019). Ownership group size is one of three properties that do not match their references, together with the in-degree tail (exponent 3.5 to 4.5, compared with 1.43 on Rabobank accounts) and per-account monthly activity (KS 0.437). We report all three as limitations.

Label sensitivity. We varied the materiality threshold τ (the minimum illicit share for a transfer to be labeled as laundering) and the spending rule (whether an account spends its illicit or its legitimate funds first). Under the released illicit-first rule, prevalence stays between 0.11% and 0.16%, and the global ranking of the detectors remains broadly similar for τ from 0.05 to 0.75 (Spearman rank correlation between 0.82 and 1.00 with the ranking at the default τ = 0.25). At τ = 0.25, switching to proportional spending reduces positive-label prevalence from 0.131% to 0.031% and the share of positives produced by propagation from 76.6% to 7.4%. The spending rule is therefore the influential choice, and we state it as a limitation.

Practitioner study. A clustered re-analysis (see Q4) gives a provenance accuracy of 0.458 (95% CI [0.283, 0.658]), which does not establish indistinguishability. A measure that needs no scorer, the share of judgments on suspicious cases that the practitioner did not dismiss (escalation answered yes or not enough information, both of which lead a bank to escalate the case), rises with context: 55.6% with transactions only (T, 10/18), 64.7% with the customer file added (TC, 11/17) and 70.6% with the investigation records added (TIC, 12/17), compared with 40.0% on legitimate controls (10/25).

Together, these results make explicit which properties of AML-view agree with external references and which do not. In the revised paper, we have added a paragraph "Fidelity beyond calibration targets" to Section 4.2, with the full checks in Appendix C (Table 8). We now describe τ and the spending rule in Section 3.2, with the sensitivity analysis in Appendix I (Table 45), report case selection, the non-blinded scorer and the clustered re-analysis in Section 5 and Appendix J.6, and list the three mismatched properties as limitations in Section 6.

AuthorsOpenReview ↗

Response to Reviewer r9Gz (2/8): W2, overlap with existing releases and importance of the added features

We appreciate the reviewer's recognition of the richer investigation data, and we agree that the original novelty claim was too broad. AML-view builds on existing work: it extends the flow tracking of AMLWorld (Altman et al., 2023), and AMLGentex already provides per-institution transaction networks (Östman et al., 2025). We have narrowed the claim in Section 1 to: "To our knowledge, no existing synthetic AML release derives institution-specific views of one shared world in which the observing institution holds customer records for its own clients and only transactional and public information for external counterparties."

AML-view also differs from AMLWorld and AMLGentex in what the world contains, not only in what a bank's view removes. AMLWorld simulates persons, firms and banks, but most of these entities are not exposed in the released data: Table 1 marks no person, firm, household, ownership, customer-profile or PEP information for it. AMLGentex assigns an age and an income to each account, but no person entity links several accounts, and no KYC record is released (Table 6). AML-view releases the records an investigation uses: persons and households, firms with their ownership links and beneficial owners, and, for each party, a KYC profile with a risk rating, PEP status, adverse-media hits and identity verification. These parties are calibrated against French population and firm statistics (35 checks, Section 4.2), and the laundering activity comes from 38 documented scenarios, each traced to its source (Section 3.2). AML-view therefore releases all 13 investigation axes, against 5 for AMLWorld and 5 for AMLGentex (Table 5). The local view then decides which of these records each bank holds: full records for its own customers, and only transfers and public registry facts for external counterparties.

AML-view thus contributes three capabilities that, taken together, no evaluated release provides:

  • A shared financial world. Every institution view is a projection of one world of persons, firms, ownership links, accounts and payments, so global and local results can be compared on the same events (Figure 3a).
  • Asymmetric institution-specific observability. A bank holds full records on its own customers and sees external counterparties only through their transfers and public registry facts (Section 3.4).
  • Investigation records linked to transaction flows. KYC, beneficial ownership, PEP status, adverse media and identity verification are attached to the parties behind each transfer (Section 3.3).
AuthorsOpenReview ↗

Response to Reviewer r9Gz (3/8): W2, how much each capability matters

We agree that the submission did not establish how much this additional coverage matters. We therefore added new experiments that measure the value of each capability.

Investigation records. We ran a per-block ablation with Investigator, our supervised reference (a gradient-boosted model on the features of the two endpoint accounts and of the transaction), in the causal regime over five seeds. Starting from transaction features only (0.514 AUROC), behavior, KYC and ownership each add signal (+0.098, +0.083 and +0.048 AUROC), while identity adds almost nothing (+0.001). All blocks together reach 0.629, a gain of 0.115 (Section 4.3; Appendix H.6, Table 43).

Asymmetric observability. We decomposed the observation gap, defined as global minus institution-local AUROC, using an intermediate view that keeps the bank-visible transactions but the global attributes of every account (worlds of 1M transactions, original prevalence, causal regime, five seeds). For Investigator, unseen transactions account for 0.050 of the 0.056 gap and withheld customer records for an additional 0.006 (Appendix H.4, Table 41). Applying the same operator to AMLWorld reproduces the edge effect, with Investigator dropping from 0.897 to 0.833 (Appendix H.5, Table 42), but the attribute part can be measured only on AML-view.

Investigation by practitioners. Pattern recovery, the share of judgments on suspicious cases in which the written answer identifies the specific laundering behavior, rises from 50.0% (9/18 judgments) with transactions only to 68.8% (11/16) with the customer file and 76.5% (13/17) with the investigation records (Section 5).

Thus, the added records carry detection signal, and among the evaluated releases the cost of withheld customer records can be measured only on AML-view. For detection, this cost is within seed-to-seed variation, whereas investigation is sensitive to these records, a finding that rests on a small practitioner panel. In the revised paper, we report the coverage comparison in Section 4.1, the decomposition and the per-block ablation in Section 4.3, and the detailed protocols and results in Appendix H.4 to H.6.

Reviewer_Y51LOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AML-view, a synthetic dataset and configurable generator for anti-money-laundering (AML) investigation under realistic, institution-specific information boundaries. It constructs a shared financial world with rich customer, ownership, due-diligence, and transaction information, instantiates 38 laundering scenarios, and derives matched institution-local views in which information about external counterparties is restricted. Experiments show that local observability generally makes AML detection harder and can change model rankings, while statistical and practitioner evaluations support the dataset’s realism and investigative value.

Critical strengths
  1. The problem is well motivated and practically important. Existing synthetic AML benchmarks assume global visibility, and AML-view fills a real gap by explicitly modeling the information boundaries that institutions actually face.
  2. The evaluation is relatively broad, covering multiple dimensions of dataset quality.
Critical weaknesses
  1. The novelty claim is imprecise and not substantiated against the closest prior work.The paper claims that "no existing synthetic AML release provides multiple institution-specific views of one common financial world", yet Table 1 marks several related generators as offering “partial support”. Following, there is no direct empirical comparison with these generators. The authors could apply the same detection models and evaluation protocol to matched global and local views constructed from AMLGentex (Östman et al., 2025) and AMLWorld (Altman et al., 2023). Without this, it is unclear whether the observed global-to-local difficulty gap is specific to AML-view’s design or would appear with any reasonable view restriction on existing data.
  2. The downstream value of the rich investigation context is not comprehensively demonstrated. Table 3 demonstrates the graph-construction ablation produces mixed results across models. Further, it does not isolate the contributions of KYC, beneficial ownership, PEP status, adverse media, and identity-verification attributes.
Questions for authors
  1. The presentation could be improved by introducing an end-to-end running example earlier in the paper. E.g, add one laundering chain in the global world and then illustrate exactly which entities, transactions, and attributes are visible to two different institutions.
  2. Table 1 should include explicit explanations for “partial support”. A minor point: some experimental results would also be easier to interpret using figures rather than tables.
  3. A large proportion of positive labels is produced through propagation, a sensitivity analysis on the propagation threshold and spending assumptions would further strengthen the evaluation.
5 comments
AuthorsOpenReview ↗

Response to Reviewer Y51L (1/5): W1, novelty claim and decomposition of the observation gap

We thank the reviewer for the constructive comments. Unless stated otherwise, page and line numbers refer to the revised PDF.

We agree that the original novelty claim (p. 2 of the submission) was too broad. We have narrowed it in Section 1 (p. 2): to our knowledge, no existing synthetic AML release derives institution-specific views of one shared world in which the observing institution holds customer records for its own clients and only transactional and public information for external counterparties. The related work (p. 3) and Table 1 now make the same distinction, and a new appendix table (Table 6) justifies every partial entry of Table 1.

1. How AML-view differs from AMLWorld and AMLGentex. AMLWorld provides sending and receiving bank identifiers, which allow transactions to be partitioned by institution. AMLGentex supports institution-specific transaction networks. AML-view also derives bank views from one shared world, but it distinguishes which transfers a bank sees from which records it can access about the parties to those transfers. For example, both banks see a cross-bank payment they process, but each sees the private customer file only for its own customer. For the external counterparty, each bank sees the transaction information and the available public registry records (Section 3.4, p. 6; Appendix D.4). This is why Table 1 distinguishes transaction-level partitioning from AML-view's institution-specific access to customer records.

2. Decomposition of the observation gap on AML-view. We use three views of each generated world: the full global view, an intermediate view that keeps only the bank-visible transactions but the global attributes of every account, and the fully local view that also masks the private records of external counterparties. The edge part is global minus intermediate AUROC, the cost of the transfers the bank never records. The attribute part is intermediate minus local AUROC, the cost of the private records it does not hold; the two views compared in this part contain the same transactions. A positive value means that AUROC is lower in the more restricted view. Worlds of 1M transactions, original laundering prevalence, causal (time-respecting) features, five seeds (Appendix H.4, Table 41):

MethodGlobalLocalΔobsEdge partAttribute part
Investigator0.6820.626+0.056+0.050+0.006
HGT0.5710.538+0.033+0.039−0.006
GCN0.5310.545−0.014−0.016+0.002
RustGraph0.5030.530−0.027−0.030+0.003

Investigator is our supervised reference, a gradient-boosted model on the features of the two endpoint accounts and of the transaction. For Investigator, restricting transactions accounts for 0.050 of the 0.056 AUROC drop, and masking private counterparty records adds 0.006. For every detector, the edge part dominates. The attribute part is within the seed-to-seed standard deviation, so it is small for detection, which the paper states. This is the distinction AML-view makes measurable: the two restrictions can be assessed separately on one fixed world.

AuthorsOpenReview ↗

Response to Reviewer Y51L (2/5): W1, the same operator on AMLWorld and AMLGentex

3. Can transaction restriction alone produce a similar gap? To answer this directly, we applied the same observation operator, detectors and protocol to AMLWorld and AMLGentex (AUROC, causal regime, five seeds; Appendix H.5, Table 42). Where possible, the observing institution holds about the same share of accounts as the observing bank in AML-view (14%). AMLWorld spreads its accounts over 30,470 very small banks, so the observer merges the 69 largest, which hold 14% of the accounts and record 39% of the transactions. The experiments shipped with AMLGentex place every account in a single bank, which makes any local view identical to the global one; we therefore ran the AMLGentex generator unchanged with 20,000 accounts over five banks of AML-view-like sizes, and the observer is the largest bank (34% of the accounts, 35% of the transactions). Neither release restricts customer attributes by bank, so on both the local view removes transactions only.

MethodAMLWorld globalAMLWorld localAMLGentex globalAMLGentex local
Investigator0.8970.8330.9410.989
RustGraph0.4290.4070.5420.831
GCN0.5010.7080.5100.937
GAT0.4980.6150.5020.915
GraphSAGE0.4590.4610.6140.563
R-GCN0.5010.4540.4980.525
HGT0.4590.4520.7040.531

On AMLWorld, Investigator's AUROC falls from 0.897 to 0.833, a drop of 0.064, similar in magnitude to the 0.050 edge part on AML-view. A drop of this size is therefore not unique to AML-view and does not require masking private counterparty records: the cost of unseen transfers is generic. On AMLGentex, the results are mixed: Investigator improves from 0.941 to 0.989, while HGT drops from 0.704 to 0.531. Restricting the transactions to one bank changes which transactions are used to train and evaluate a detector; in particular, it raises the share of laundering transactions (0.026% locally against 0.018% globally) without reproducing an institution boundary. Transaction restriction can therefore change detector performance, but the direction is not universal.

What AML-view adds is the attribute restriction, which neither release can express. It is the first release on which the detection gap of one bank can be split into the part due to unseen transfers and the part due to withheld customer records. Section 4.3 summarizes this comparison (p. 8), and Section 1 points to it (p. 2).

4. Where the private records matter. AML-view also links transactions to customer, ownership and due-diligence records, so it can measure what this evidence adds once a case is under investigation. In the practitioner study (Section 5, p. 10), the pattern-recovery rate rose from 0.500 (9/18 judgments) with transactions alone to 0.688 (11/16) with the customer file and 0.765 (13/17) with the full investigation layer. The detector experiment varies access to private counterparty attributes and measures transaction ranking; the practitioner study varies the evidence in a case file and measures whether investigators identify the underlying pattern. AML-view supports both questions on the same data model.

AuthorsOpenReview ↗

Response to Reviewer Y51L (3/5): W2, value of the investigation context

We agree that the graph-construction ablation (Table 3 of the submission, now Figure 3b and Table 36) changes how the graph is built and does not isolate the contribution of each attribute block. We added a separate block-level ablation for supervised detection, and we explain below why the numbers are the way they are.

Contribution of each block to detection. Starting from transaction features alone, we added one due-diligence block at a time to Investigator (causal regime, five seeds; Appendix H.6, Table 43):

FeaturesAUROCGain over transactions only
Transactions only0.514
+ behavior+0.098
+ KYC+0.083
+ ownership+0.048
+ identity+0.001
All blocks0.629+0.115

Behavior gives the largest individual gain, followed by KYC and ownership, while the identity block adds almost nothing on its own. All blocks together give the best result, 0.629 AUROC, 0.115 above transactions alone. This joint gain is smaller than the sum of the individual gains, which is consistent with some blocks carrying overlapping information. These records therefore carry detection signal that a transaction-only release cannot expose. Section 4.3 reports these results in the representation-ablation paragraph (p. 9).

Why the graph-construction results are mixed. Adding identity edges changes which accounts each graph detector sees as neighbors, rather than adding fields to a transaction row. The detectors of Figure 3b are unsupervised and score structural anomalies. Identity edges add dense clusters around shared identifiers, which raises AUROC by 0.072 for RustGraph but changes the neighborhoods that GAT aggregates over and lowers its AUROC by 0.070. Without labels, these detectors cannot learn which added links are informative, whereas a supervised model can, as the block ablation shows. The value of node attributes also depends on what the transaction features already contain. In the static regime, removing the full node-attribute block lowers the precision of Investigator at the 1% alert budget from 0.801 to 0.071 (Table 38). In the causal regime, the effect is much smaller, because pre-transaction behavioral features remain available on the edges.

Value for case investigation. In Section 5, practitioners identified the specific laundering pattern in 9/18 judgments (0.500) with transactions alone, 11/16 (0.688) with the customer file added, and 13/17 (0.765) with the full investigation layer (p. 10). The block ablation measures what each block contributes to transaction ranking, while the practitioner study shows how progressively richer case files support the reconstruction of the laundering pattern.

Reviewer_fA4NOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces AML-view, a synthetic AML dataset and generator that models a shared financial world of persons, firms, ownership, accounts, due-diligence attributes, and payments, instantiating 38 documented laundering scenarios and propagating illicit value through downstream transfers. Each world contains about 10 million transactions over 70,000 accounts, and an institution's local view retains private records for its own customers, keeping mainly transactional and public information for external counterparties, so differences between views reflect the observation boundary.

Critical strengths

1-** Important problem:** The paper addresses a practical problem with real-world implications for detecting money laundering. Better AML datasets could help develop more effective detection systems.

2- One-world, multiple-view design: The paper introduces an interesting design where multiple institutions observe different parts of the same financial world. The approach is well positioned, with additional implementation details in the appendix.

Critical weaknesses

1- Presentation is very difficult to follow: The paper introduces many domain-specific terms and acronyms without enough explanation for readers unfamiliar with AML. Terms such as TVD and FATF could be introduced more clearly, and Table 1 is difficult to follow because of abbreviations such as "Pers." and "Cust.".

2- Overstated contribution language: The contributions are headed "Holistic, statistically grounded financial-world simulation" and "Rich, document-grounded laundering scenarios" (l.069, l.074), and the abstract calls the world "investigation-rich" (l.016), but the evaluation supports a narrower claim than the wording implies.

3- Baselines uncited at first use: TADDY appears in the intro carrying a headline number ("falls from 0.739 to 0.510 AUROC") but is neither cited nor described until Table 2.

4- Table 1 legend is incomplete and hard to understand : the caption glosses HH, Own., Inst.-local obs. and Ground. but leaves Pers., Cust. profile, PEP/media/identity, AML label and Typ. label unexplained. table 1 is hard to follow like Pers. ?

5- No worked example: Table 11 lists fields under global vs. local visibility, but the paper never shows a single example record, transaction or local-view case, so a reader cannot see what the release actually contains.

6- Baseline coverage to thin: The experiments include several graph-based detectors, but do not evaluate recent tabular foundation models such as TabPFN, TabICL, or Mitra. Including these models on the structured transaction and customer features would help establish how useful the dataset is beyond graph-based detection.

Questions for authors

Q1- Synthetic data fidelity: Since the dataset is entirely synthetic, how do the authors validate that the generated transactions and financial networks resemble real banking activity beyond the statistics used for generation?

Q2- Demographic bias: Have the authors evaluated whether synthetic demographic, KYC, and customer risk attributes introduce or amplify biases in downstream AML detection?

Q3- Additional baselines: How would agentic approaches or tabular foundation models such as TabPFN, TabICL, and Mitra perform on this dataset?

Q4- Oracle performance: What is the best achievable detection performance (upper bound) using the full global information and ground-truth labels? How large is the gap between this oracle reference and the evaluated models?

5 comments
AuthorsOpenReview ↗

Response to Reviewer fA4N (1/5): W1 to W3

We thank the reviewer for the careful reading. Unless stated otherwise, page and line numbers refer to the revised PDF.

W1. Acronyms and Table 1. Thank you for pointing this out. The submission explained both terms, but only after their first use in the introduction: TVD was used on p. 2 and defined in Appendix C.1 (p. 15), and FATF was used on p. 2 and expanded only in citation form in Section 3.2 (p. 4). We agree that readers should not have to look ahead for these explanations. We have made the following changes:

  • Every acronym is now expanded at first use in the main text: total variation distance (TVD, p. 2), Financial Action Task Force (FATF, the intergovernmental body whose 40 Recommendations define AML obligations, p. 2), know your customer (KYC), politically exposed person (PEP), suspicious activity report (SAR) and financial intelligence unit (FIU).
  • Section 2 now opens with a three-sentence AML primer that introduces customer due diligence, KYC, PEP screening, SARs and FIUs (p. 3).
  • A glossary (Appendix M, Table 50) defines every acronym used in the paper. For example, TVD is defined there as the share of probability mass that must move to turn the generated distribution into the reference one.
  • All column names of Table 1 are written in full (see W4).

W2. Contribution wording. Thank you for pointing out where the contribution headings could be more precise. We have clarified what is integrated, what is statistically calibrated, and what is grounded in published documents (p. 2).

Financial-world simulation. By "holistic", we mean that persons, households, firms, ownership links, accounts and payments are represented in one connected financial world. To make the scope of the statistical grounding explicit, the heading is now "Holistic financial-world simulation with population calibration." The 35 checks measure agreement with the reference distributions used to construct the population (Section 4.2, Table 9), but do not validate payment behavior or ownership dynamics. The contribution now states this distinction (p. 2), alongside the existing statement in Section 3.1 that laundering prevalence, typology mixtures and some structural parameters remain experimental choices (p. 4), and the list of quantities that are not calibrated in Appendix F.5. Properties that the generator never fits are now tested separately (Appendix C; see Q1).

Laundering scenarios. The second heading is now "Document-grounded laundering scenarios and flow-based labels." Each of the 38 laundering scenarios is traceable to a published source, and 16 are based on complex cases described in reports by the FATF and three FIUs (Section 3.2, p. 5). This establishes the documentary basis of the catalog, not the completeness of the typology space, which the contribution now states (p. 2). The simulator then separately tracks illicit value through the ledger to label transactions, including downstream transfers not directly generated by a scenario (Appendix D.3).

Abstract. We have replaced "investigation-rich" with "investigation-oriented" (p. 1) to emphasize the purpose of the linked customer, ownership and due-diligence records without making a broader claim about their coverage.

W3. Baselines cited at first use. TADDY and RustGraph were cited in Section 4.3 and in the results tables, where DyGADBench was also introduced, but the introduction reported TADDY's result before either citation. Both detectors are now cited at their first mention and described in one clause each, and DyGADBench is identified as the source of their implementations (p. 2). Every method named in Sections 1 and 2 now carries its citation at first use.

AuthorsOpenReview ↗

Response to Reviewer fA4N (2/5): W4 and W5

W4. Table 1 legend. We appreciate this detailed check of Table 1, and we have revised it so that it can be read on its own (p. 4).

  • Every column name is now spelled out in the header: Person, Firm, Household, Ownership, Customer profile, PEP/media/identity, Local view, AML label, Typology label and Grounding.
  • The caption defines each column that is not self-explanatory: Customer profile (the KYC record: occupation, income band, risk rating), PEP/media/identity (politically exposed person status, adverse-media hits, identity verification), Local view (an institution-local projection with restricted information on external counterparties), Typology label (the laundering scenario that produced the transaction) and Grounding (calibration against public statistics or documents).

We have also added an appendix table (Table 6, p. 17) that justifies every partial entry (△) of Table 1 in one sentence, stating what each release contains and what it lacks. For example, AMLWorld is partial for the local view because each transaction records the sending and receiving bank, so a single bank's view can be obtained by filtering, but no customer attribute is restricted by bank. AMLGentex is partial for persons and customer profiles because each account is assigned an age and an income drawn from Statistics Sweden, but these attributes sit on accounts, and no person entity links several accounts.

W5. Worked example. In the submission, Figure 2 (p. 4) followed one laundering chain (proceeds source p, mule accounts m1 and m2, sink s) as it exists in the world and as banks A and B record it; Appendix C.4 and Table 11 (p. 18) specified which fields each view keeps; and a compact copy of the seed-44 instance was in the supplementary archive. That example was schematic: it showed which records exist, not what they contain. We agree that a reader should be able to see the release content directly.

  • We have replaced Figure 2 (p. 5) with a record-level version of one released seed-44 chain, shown side by side: (a) the global world G, with the parties' KYC fields, the beneficial-ownership link, and each transfer's amount and timestamp; (b) what bank A records, with full customer records for its own customers and only the account identifier and public registry fields (firm, NAF code, ownership) for the external counterparty; (c) what bank B records, with the roles reversed.
  • A new appendix table (Table 12, p. 22) shows one complete record of each released table from this chain (person and firm with their KYC fields, account, ownership, transaction with its illicit amount and laundering label, and provenance), so that a reader sees the release content without downloading it. Appendix D.4 and Table 11 (p. 21) specify the visibility tiers.
AuthorsOpenReview ↗

Response to Reviewer fA4N (3/5): W6 and Q3, tabular foundation models and agentic approaches

Why the benchmark centers on graph detectors. Many laundering patterns span several accounts, so the relevant signal may be absent from individual transaction records or account histories (Altman et al., 2023). Work on real bank data has shown the value of modeling links between customers (Johannessen & Jullum, 2025). Revolut's PRAGMA, which encodes event histories in isolation, obtains a 47.1% lower F0.5 than its network-aware AML production baseline (Ostroukhov et al., 2026). On four smaller AMLWorld datasets, adding a GNN to representations from a time-series foundation model improved PR-AUC by a further 4.8% (Savvidi et al., 2026). Tabular models can also benefit from relational information when it is supplied as subgraph features (Blanuša et al., 2024). This rationale is now stated in Section 4.3 (p. 7) and Appendix H.1.

The submission already included a row-wise tabular baseline: Investigator, a gradient-boosted model (GBM) over transaction features and the features of the source and destination accounts. It reaches 0.893 global and 0.831 institution-local AUROC (Figure 3a, Table 34). It is an optimistic supervised reference, trained with laundering over-represented on simulator ground-truth labels, whereas the other detectors in Figure 3a are unsupervised or weakly supervised.

New results. Following the reviewer's suggestion, we evaluated TabPFN v2, TabICL and Mitra, each with and without the subgraph features of the Graph Feature Preprocessor. Although pretrained, these models are not label-free AML detectors: they predict by in-context learning over labeled training rows. We therefore evaluated them with GBM in the supervised setting, at the original laundering prevalence (about 0.1%, where a random score has an AUPRC of about 0.001), on five worlds of 200,000 transactions, in the global and institution-local views and in both feature regimes. Global view, static regime:

ModelAUROCAUROC + subgraphAUPRCAUPRC + subgraph
GBM (Investigator)0.6410.6670.0180.030
TabPFN v20.6600.6410.0260.027
TabICL0.6860.6900.0430.061
Mitra0.6380.6400.0190.024

All four models reach 0.64 to 0.69 AUROC, and TabICL is highest (0.686, 0.045 above GBM); with few positives per test split, we do not claim that any model is clearly superior. All four lose AUROC in the institution-local view, by 0.018 to 0.139 in the causal regime and 0.055 to 0.187 in the static regime, so the observation boundary affects row-wise models as well as graph detectors. Subgraph features raise global AUPRC for all four models in both regimes, while their effect on AUROC is mixed. Graph information can thus improve tabular detectors even when they do not operate on a graph. These models still receive reliable simulator labels that are generally unavailable in routine AML monitoring, so this experiment does not establish performance when labels are scarce.

In the revised paper, these models appear in Figure 3a (rows marked †), in Section 4.3 (pp. 8 to 9), and in Appendix H (Table 35, causal regime in both views; Table 37, both regimes with and without subgraph features).

Q3, agentic approaches. Evaluating agents is beyond the scope of this paper, but AML-view provides the case files and scoring protocol needed to benchmark agents on case-level investigation. Recent agentic AML work addresses single steps of the workflow, such as drafting SAR narratives (Naik et al., 2025) and adverse-media screening (Chernakov et al., 2025). AML-view's institution-local case files come at three information levels: transactions only (T), plus the customer file (TC), and plus the investigation records (TIC) (Section 5, p. 9). Using the same cases, conditions and pattern-recovery criterion, agents can be compared with human investigators, and the benefit of each additional evidence layer can be measured. The conclusion lists this comparison as future work (p. 10).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.