acceptodds
Under review as a conference paper at ICLR 2027

LEAP-HGN: A Brain-Inspired Amygdala-Parietal Detection Framework with LLM-Driven Evolutionary Rules for Anti-Money Laundering

Abstract

Money laundering activities are increasingly evolving into a time-sensitive, dense micro-circulation pattern, posing severe challenges to existing graph neural network technologies in dealing with the disruption of continuity caused by discrete time slices, the signal drowning due to massive business noise, and the logical lag resulting from the dynamic evolution of strategies. To address these issues, we propose the LEAP-HGN (LLM-Evolved Amygdala-Parietal HyperGraph Network), a novel approach that simulates the human brain's threat processing mechanism within a unified brain-like cognitive framework. This model first constructs a spatio-temporal multi-view money laundering hypergraph, explicitly restoring the disrupted temporal causal chains of funds. On this basis, we redesign a biomimetic amygdala-parietal synergy mechanism: the amygdala low pathway uses neural gain modulation technology to transform expert rules into explicit warning factors, forcibly amplifying the response of sparse high-risk nodes before deep feature propagation, thereby preventing signal dilution from the source; the parietal high pathway performs deep spatial reasoning through Markov high-order diffusion and residual attention methods, penetrating the dense noise background from a global perspective and reconstructing hidden long-distance multi-hop links. Additionally, we design a Large Language Model (LLM) self-evolutionary closed loop based on hard sample mining, an LLM-driven loop that parses hard samples into symbolic rules for autonomous re-injection, achieving continuous self-evolution without human intervention. Extensive experiments on six money laundering datasets demonstrate that LEAP-HGN significantly outperforms 12 SOTA baselines. Code and datasets will be made publicly available at https://anonymous.4open.science/r/LEAP-9AEF upon acceptance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Financial fraud detection, Graph learning, Anomaly detection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tRgvOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes LEAP-HGN, a brain-inspired neuro-symbolic framework for anti-money laundering detection. It combines spatio-temporal multi-view hypergraphs, rule-guided risk amplification, and high-order structural reasoning to identify hidden laundering patterns. An LLM-based feedback mechanism further updates symbolic rules using hard samples. The method is evaluated on six financial datasets against 12 baselines, with strong results across most evaluation metrics.

Critical strengths

(1) This is the first work to introduce brain-like intelligent perception mechanisms to the field of anti-money laundering detection. This distinctive research perspective offers substantial originality and and novelty, and the addressed research issues have significant practical application value. Specifically, by integrating amygdala-driven rapid risk amplification and parietal-guided global spatial reasoning, this study unifies local risk intuition with global structural perception to better identify concealed syndicates.

(2)This study designs a novel Markov structural diffusion mechanism deployed over node‑hyperedge‑node paths to capture high‑order structural information, which successfully bridges time‑slicing causal breaks and explicitly reconstructs global long‑range fund dependencies. Additionally, they establish a self‑evolutionary rule closed‑loop via LLM hard sample mining to overcome the rigidity of static controls against dynamic laundering tactics. This enables autonomous neuro‑symbolic evolution, allowing the model to continuously adapt to evolving fraud patterns without manual intervention. The experimental design is also very thorough. The validation on multiple datasets proves that the proposed model has excellent performance.

Critical weaknesses

None

Questions for authors

(1) In Section 3.4, the term "hard negative samples" seems slightly inconsistent with Equation (12), which describes positive samples missed by the model. The intended meaning is clear from the equation, but the terminology could be more precise. (2) There is a small naming inconsistency between "Account-Bank path" in Section 3.1 and "Account-Bank View" in Figure 1. Using consistent terminology would make the description easier to follow. (3) In Section 4.4, the phrase "optimal configuration" could be more specific about whether it refers to the parameter settings examined in the experiments. (4) Table 1 reports results as percentages, whereas Figures 2 and 3 use decimal values. A brief clarification in the figure captions would make the results easier to compare.

Reviewer_9ZhPOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LEAP-HGN, a neuro-symbolic framework for account-level anti-money-laundering detection. It combines a four-view transaction hypergraph, rule-derived risk amplification, residual attention and high-order diffusion, and an LLM-driven loop that generates corrective rules from training false negatives. Evaluation covers six financial datasets, twelve baselines, component ablations, and analyses of rule-evolution trajectories under a chronological out-of-time protocol.

Critical strengths

The framework translates summaries of missed positive samples into constrained symbolic rules, screens these candidates, and incorporates accepted rules into subsequent neural training. This is a potentially useful approach to updating expert-informed detectors without invoking an LLM for every transaction.

Critical weaknesses

ARC and EdgePrompt are categorized as LLM-based methods, although their original formulations concern graph anomaly detection and graph prompt tuning, respectively. Fraud-R1 is primarily a benchmark for LLM robustness against fraud and phishing inducements, rather than a directly applicable account-level transaction classifier. The manuscript does not adequately explain how these methods were implemented and adapted to produce the reported results. Together with unclear transaction-to-account mappings across datasets, this prevents a reliable assessment of experimental comparability.

How were transaction-level datasets converted into account-level graphs and labels?

What existing evidence isolates the benefit of LLM-generated rules from additional training, an expanding rule set, or conventional threshold search? Comparisons matching the candidate-rule budget and neural training budget would materially affect my assessment of the LLM component.

The “brain-inspired” framing is not sufficiently substantiated by the methodology or experiments. The Amygdala and Parietal modules are implemented primarily through rule-derived risk scores, feature amplification, residual attention, and graph diffusion. Although these operations may be useful for anti-money-laundering detection, the manuscript does not establish what specific computational constraints or predictions are introduced by the proposed correspondence to brain regions. The current ablations assess whether individual modules improve performance, but do not demonstrate that their biological organization provides an advantage over a conventional rule-guided graph-learning architecture. Consequently, the neuroscience terminology currently contributes more to the presentation than to the technical justification. The authors should either identify and validate the design principles uniquely motivated by this analogy or present the method more directly as a neuro-symbolic graph-learning framework.

Questions for authors

See the weakenss.

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

The paper proposes LEAP-HGN for account-level anti-money laundering detection. It combines a four-view hypergraph, rule-guided risk amplification, residual attention, and graph diffusion. Its Self-Evolutionary Rule Closed-Loop (SRCL) uses an LLM to generate symbolic rules from aggregated features of training-set false negatives. Experiments on six datasets against twelve baselines report the best results on most metrics.

Critical strengths
  1. The problem is practically important. Combining interpretable rule priors with graph learning and refining them using missed laundering cases is a worthwhile direction.

  2. The chronological out-of-time evaluation and restriction of laundering templates to training positives are sensible measures against temporal leakage.

Critical weaknesses
  1. Limited technical novelty beyond integrating existing components. Multi-view hypergraphs, rule-guided attention modulation, residual attention, and graph diffusion build on established techniques. The paper does not clearly identify a new mechanism arising from their integration, and the component-removal ablations do not establish the distinctive value of the proposed interactions. LLM-based rule refinement is the more distinctive element, but it lacks comparison with non-LLM search over the same rule space. The contribution therefore appears primarily incremental.

  2. The adaptation claims exceed what the method and experiments demonstrate. SRCL refines rules using training-set errors and screens them on historical data. The reported evaluation tests generalization to later data, but does not establish adaptation after new laundering strategies emerge. The paper also insufficiently discusses continual learning for AML. Separately, the within-window hypergraph construction does not clearly explain how transaction chains fragmented across window boundaries are recovered.

  3. Baseline comparisons require clarification. ARC and EdgePrompt are categorized as LLM-based methods, although their original formulations concern graph anomaly detection and GNN prompt tuning, respectively. Fraud-R1 is an evaluation benchmark rather than a standalone account detector. Their implementation under the paper's account-level task is insufficiently described. The discussion also emphasizes selected baselines: the 20.88-point XploreML AUC gain is over HOGRL, whereas the gain over the strongest baseline, ARC, is 0.70 points.

  4. Reporting is insufficient to assess reliability. Ablation and per-cycle results are largely presented as plots without exact numerical tables. Independent run counts, seeds, and numerical variability are not reported. The basis for selecting Cycle 7 and constructing the confidence band in Appendix M is unclear. Section 4.4 also claims stability with respect to λ_alert, but Figure 3 does not plot that parameter.

Questions for authors
  1. Can you demonstrate adaptation through rolling temporal updates or a held-out laundering-typology experiment? How does the method preserve transaction chains across window boundaries?

  2. How does LLM rule generation compare with non-LLM search using the same features, rule space, and screening criteria, and with adding those features directly to the model?

  3. Please clarify the implementation and categorization of ARC, EdgePrompt, and Fraud-R1, and evaluate Multi-GNN/PNA under the same account labels, temporal splits, and inputs.

  4. Please provide numerical ablations, results before the first LLM-generated rule, and statistics across independent runs. Was Cycle 7 selected on validation data? How was the confidence band computed, and what happens when λ_alert = 0?

1 comment

We sincerely thank the reviewer again for the constructive feedback. We hope our responses and revisions have adequately addressed the concerns. Please let us know if there are any remaining questions or concerns. Thank you again for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.