Kailing Fu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Latest Publications
- Open Access
- 22 October 2026
- Published by MDPI AG in Preprints.org
Early Credit-Card Fraud Detection via Behavior-Deviation Sequence Modeling
Credit-card fraud detection protects the payment networks through which issuers, merchants, and cardholders transact, and it must decide under a fixed alert budget which transactions to stop. Sequence models capture cardholder behavior more faithfully than models that score each transaction in isolation, but most prior works evaluated them with aggregate metrics that treat every fraudulent transaction of a compromised card as equally valuable, hiding how early in a fraud burst the detector fires, which is what determines the realized loss. We propose DevSeq, a behavior-deviation sequence model for early card-fraud detection that explicitly represents how far a transaction departs from the cardholder's own habits. DevSeq first encodes the recent transaction sequence of a card with a Transformer encoder and summarizes the long-term habits of the same card with a causal behavioral profile, then derives an explicit deviation representation by cross-attending the current transaction to the recent sequence and by comparing it against the projected long-term profile, and finally adds a self-supervised next-transaction prediction task that requires no fraud labels. We further propose a burst-level evaluation protocol that, under a fixed alert budget, separates the first fraudulent transaction of a fraud burst from the subsequent ones and reports the loss incurred before the first alert. On a public synthetic dataset of 24 million card transactions under a temporal split, DevSeq raises first-transaction recall at a 0.5% alert budget from 0.182 for an aggregation-feature baseline and 0.268 for the strongest sequence baseline to 0.342, and reduces the mean loss before the first alert from 199.1 to 152.8 USD. With only 10% of the fraud labels, DevSeq still exceeds the aggregation baseline trained with all of them.
- 22 October 2026
- Published by MDPI AG in Preprints.org
Open AccessEarly Credit-Card Fraud Detection via Behavior-Deviation Sequence Modeling
Credit-card fraud detection protects the payment networks through which issuers, merchants, and cardholders transact, and it must decide under a fixed alert budget which transactions to stop. Sequence models capture cardholder behavior more faithfully than models that score each transaction in isolation, but most prior works evaluated them with aggregate metrics that treat every fraudulent transaction of a compromised card as equally valuable, hiding how early in a fraud burst the detector fires, which is what determines the realized loss. We propose DevSeq, a behavior-deviation sequence model for early card-fraud detection that explicitly represents how far a transaction departs from the cardholder's own habits. DevSeq first encodes the recent transaction sequence of a card with a Transformer encoder and summarizes the long-term habits of the same card with a causal behavioral profile, then derives an explicit deviation representation by cross-attending the current transaction to the recent sequence and by comparing it against the projected long-term profile, and finally adds a self-supervised next-transaction prediction task that requires no fraud labels. We further propose a burst-level evaluation protocol that, under a fixed alert budget, separates the first fraudulent transaction of a fraud burst from the subsequent ones and reports the loss incurred before the first alert. On a public synthetic dataset of 24 million card transactions under a temporal split, DevSeq raises first-transaction recall at a 0.5% alert budget from 0.182 for an aggregation-feature baseline and 0.268 for the strongest sequence baseline to 0.342, and reduces the mean loss before the first alert from 199.1 to 152.8 USD. With only 10% of the fraud labels, DevSeq still exceeds the aggregation baseline trained with all of them.
- Open Access
- 20 October 2026
- Published by MDPI AG in Preprints.org
Knowing When to Ask: Multi-Signal Calibrated Escalation for LLM Fraud-Investigation Agents
Fraud-investigation agents built on large language models are deployed under selective autonomy, closing the routine cases themselves and escalating the difficult ones, because a compliance function cannot accept full automation. That decision is commonly driven by the confidence the agent states about itself, which language models overstate and which a reviewer may accept more readily for coming with a fluent rationale. We build CEG, a calibrated escalation gate over four signals, and use it to measure which of them actually predict when the agent is wrong: stated confidence, self-consistency across repeated independent runs, disagreement with a conventional tabular fraud classifier, and the sufficiency of the evidence retrieved. A calibrator trained on a small labeled development set maps them to the probability that the agent's conclusion is correct, and an asymmetric cost-sensitive rule escalates under an analyst capacity constraint, demanding more certainty to close a case as benign than to conclude fraud. On mobile-money simulator cases, after removing a balance shortcut identified in prior work and with the analyst modeled as an oracle of fixed capacity, three findings stand out. Stated confidence, the signal deployments commonly use, is the weakest gate that is not random, at a calibration error of 0.069 against 0.022 for the combination, and dropping it from the combination changes team cost by 2 units. Disagreement with the classifier carries the gate, and the classifier score alone therefore comes within 12% of the full combination at a fraction of its inference cost. A conformal variant holds its 10% false-negative target within the calibration period and fails it in all ten repetitions once fraud patterns shift, while the share of cases it decides automatically barely moves, so a deployment cannot detect the violation from its own coverage.
- 20 October 2026
- Published by MDPI AG in Preprints.org
Open AccessKnowing When to Ask: Multi-Signal Calibrated Escalation for LLM Fraud-Investigation Agents
Fraud-investigation agents built on large language models are deployed under selective autonomy, closing the routine cases themselves and escalating the difficult ones, because a compliance function cannot accept full automation. That decision is commonly driven by the confidence the agent states about itself, which language models overstate and which a reviewer may accept more readily for coming with a fluent rationale. We build CEG, a calibrated escalation gate over four signals, and use it to measure which of them actually predict when the agent is wrong: stated confidence, self-consistency across repeated independent runs, disagreement with a conventional tabular fraud classifier, and the sufficiency of the evidence retrieved. A calibrator trained on a small labeled development set maps them to the probability that the agent's conclusion is correct, and an asymmetric cost-sensitive rule escalates under an analyst capacity constraint, demanding more certainty to close a case as benign than to conclude fraud. On mobile-money simulator cases, after removing a balance shortcut identified in prior work and with the analyst modeled as an oracle of fixed capacity, three findings stand out. Stated confidence, the signal deployments commonly use, is the weakest gate that is not random, at a calibration error of 0.069 against 0.022 for the combination, and dropping it from the combination changes team cost by 2 units. Disagreement with the classifier carries the gate, and the classifier score alone therefore comes within 12% of the full combination at a fraction of its inference cost. A conformal variant holds its 10% false-negative target within the calibration period and fails it in all ten repetitions once fraud patterns shift, while the share of cases it decides automatically barely moves, so a deployment cannot detect the violation from its own coverage.
- Open Access
- 20 October 2026
- Published by MDPI AG in Preprints.org
Evidence-Ledger Multi-Agent Investigation for Traceable Anti-Money-Laundering Reporting
Investigating an alert and documenting the result is the most labor-intensive stage of anti-money-laundering compliance, and recent work proposed large language model agents that draft investigation narratives. A narrative that cites a nonexistent transaction or misstates an amount is a regulatory liability, yet, to our knowledge, prior investigation agents were evaluated by expert rating, detection scores or evidence recall without checking each cited transaction against the ledger. We propose ELMI, an evidence-ledger multi-agent investigation framework with a division of verification labor: an investigator agent may cite only records a tool call returned, logged in a replayable ledger; a deterministic checker verifies each cited record field by field; an independent verifier agent, which sees the claims but neither the reasoning trace nor the case label, judges whether the claimed typology is supported; and any surviving discrepancy reaches the analyst as an open issue. Our benchmark comes from a synthetic dataset whose ground-truth patterns supply each case's typology and transactions. Because the checker applies the comparison that defines citation faithfulness, we read faithfulness as a property of the pipeline rather than of the model, and rest the research claim on a single agent equipped with the same checker. Across 240 cases field faithfulness is 0.97 against 0.76 for a single tool-using agent, and it stays between 0.96 and 0.98 on three backbones and in every case-size stratum, where the single agent's falls from 0.81 on the smallest cases to 0.72 on the largest. Over the checker-equipped agent, independent verification raises typology accuracy from 0.58 to 0.67 and evidence recall from 0.59 to 0.71 at 2.6 times the tokens; at a matched step budget that agent reaches 0.60 and 0.63, leaving 0.07 and 0.08 to verification against a run-to-run standard deviation of up to 0.06.
- 20 October 2026
- Published by MDPI AG in Preprints.org
Open AccessEvidence-Ledger Multi-Agent Investigation for Traceable Anti-Money-Laundering Reporting
Investigating an alert and documenting the result is the most labor-intensive stage of anti-money-laundering compliance, and recent work proposed large language model agents that draft investigation narratives. A narrative that cites a nonexistent transaction or misstates an amount is a regulatory liability, yet, to our knowledge, prior investigation agents were evaluated by expert rating, detection scores or evidence recall without checking each cited transaction against the ledger. We propose ELMI, an evidence-ledger multi-agent investigation framework with a division of verification labor: an investigator agent may cite only records a tool call returned, logged in a replayable ledger; a deterministic checker verifies each cited record field by field; an independent verifier agent, which sees the claims but neither the reasoning trace nor the case label, judges whether the claimed typology is supported; and any surviving discrepancy reaches the analyst as an open issue. Our benchmark comes from a synthetic dataset whose ground-truth patterns supply each case's typology and transactions. Because the checker applies the comparison that defines citation faithfulness, we read faithfulness as a property of the pipeline rather than of the model, and rest the research claim on a single agent equipped with the same checker. Across 240 cases field faithfulness is 0.97 against 0.76 for a single tool-using agent, and it stays between 0.96 and 0.98 on three backbones and in every case-size stratum, where the single agent's falls from 0.81 on the smallest cases to 0.72 on the largest. Over the checker-equipped agent, independent verification raises typology accuracy from 0.58 to 0.67 and evidence recall from 0.59 to 0.71 at 2.6 times the tokens; at a matched step budget that agent reaches 0.60 and 0.63, leaving 0.07 and 0.08 to verification against a run-to-run standard deviation of up to 0.06.
- Open Access
- 20 October 2026
- Published by MDPI AG in Preprints.org
Knowing When to Hand Over: A Survey of Uncertainty, Deferral, and Escalation in Human–AI Fraud Investigation
Language-model agents are being proposed to investigate fraud and money-laundering alerts under selective autonomy: the agent closes some cases itself and escalates the rest to a human analyst. Which cases it escalates determines the division of labor, the analysts' workload, and the errors that reach customers and regulators, yet the knowledge needed to design that rule is divided among learning to defer and selective prediction, uncertainty estimation for language models, and research on human reliance on AI advice. This survey connects the three through the escalation loop of a fraud-investigation team and restates known results from these literatures in one notation. We state the cost-optimal escalation rule, which requires different confidence thresholds for closing a case as benign and for concluding fraud, and its form under a fixed analyst capacity; show that calibration suffices for selective risk to be monotone and interpretable; state the conformal bound on the rate of cases that are both closed automatically and wrong; show when combining uncertainty signals helps and when correlated signals double-count; show that the error among automatically closed cases is not identified without audits and is estimable with them; and quantify how analyst reliance on the agent's verdict erodes the value of escalation. Against this basis we compare deferral rules and uncertainty signals, and conclude that threshold-based deferral rules assume a calibrated confidence that agents do not yet provide, that agent-level uncertainty estimation is only beginning, and that, to our knowledge, no peer-reviewed study has yet evaluated an escalation rule for a fraud-investigation agent together with the analysts who receive its cases.
- 20 October 2026
- Published by MDPI AG in Preprints.org
Open AccessKnowing When to Hand Over: A Survey of Uncertainty, Deferral, and Escalation in Human–AI Fraud Investigation
Language-model agents are being proposed to investigate fraud and money-laundering alerts under selective autonomy: the agent closes some cases itself and escalates the rest to a human analyst. Which cases it escalates determines the division of labor, the analysts' workload, and the errors that reach customers and regulators, yet the knowledge needed to design that rule is divided among learning to defer and selective prediction, uncertainty estimation for language models, and research on human reliance on AI advice. This survey connects the three through the escalation loop of a fraud-investigation team and restates known results from these literatures in one notation. We state the cost-optimal escalation rule, which requires different confidence thresholds for closing a case as benign and for concluding fraud, and its form under a fixed analyst capacity; show that calibration suffices for selective risk to be monotone and interpretable; state the conformal bound on the rate of cases that are both closed automatically and wrong; show when combining uncertainty signals helps and when correlated signals double-count; show that the error among automatically closed cases is not identified without audits and is estimable with them; and quantify how analyst reliance on the agent's verdict erodes the value of escalation. Against this basis we compare deferral rules and uncertainty signals, and conclude that threshold-based deferral rules assume a calibrated confidence that agents do not yet provide, that agent-level uncertainty estimation is only beginning, and that, to our knowledge, no peer-reviewed study has yet evaluated an escalation rule for a fraud-investigation agent together with the analysts who receive its cases.
- Open Access
- 20 October 2026
- Published by MDPI AG in Preprints.org
From Scorecards to Agents: A Survey of Anti-Fraud Models Across the Model-Risk Boundary, with a Focus on Evidence Traceability
The revised United States guidance on model risk management, issued in April 2026, applies its principles to traditional statistical and quantitative models and to non-generative, non-agentic AI models, and places generative and agentic AI explicitly outside its scope. This survey follows the models used against fraud and financial crime across that line, from rules, scorecards, and statistical tests, through non-generative machine learning, to generative models and language-model agents, across six fraud types. Unlike prior surveys, we place all three generations in one decision-theoretic frame: we restate in one notation known results on cost-optimal and capacity-constrained alerting, undersampling calibration, and the identifiability of recall from alert feedback, express each generation's model, explanation, and validation quantity in that frame, and add two stylized models, one showing why mean aggregation dilutes the neighborhood signal of camouflaged fraudsters and one showing how verifier independence bounds the residual errors in a generated narrative. We find that statistical models and non-generative machine learning have mature validation practice anchored in supervisory guidance, that generative AI now appears both as a tool of fraudsters and as a component of defenses, and that in hybrid pipelines the generative component can write what an analyst or examiner reads although no dedicated supervisory framework covers it. We therefore formalize the evidence-traceability controls that can supply this missing validation as a division of verification labor, between the drafting agent, a deterministic field check, an independent verifier and the human examiner, governed by three parameters an institution can estimate; we then compare the generations quantity by quantity and close with a research agenda.
- 20 October 2026
- Published by MDPI AG in Preprints.org
Open AccessFrom Scorecards to Agents: A Survey of Anti-Fraud Models Across the Model-Risk Boundary, with a Focus on Evidence Traceability
The revised United States guidance on model risk management, issued in April 2026, applies its principles to traditional statistical and quantitative models and to non-generative, non-agentic AI models, and places generative and agentic AI explicitly outside its scope. This survey follows the models used against fraud and financial crime across that line, from rules, scorecards, and statistical tests, through non-generative machine learning, to generative models and language-model agents, across six fraud types. Unlike prior surveys, we place all three generations in one decision-theoretic frame: we restate in one notation known results on cost-optimal and capacity-constrained alerting, undersampling calibration, and the identifiability of recall from alert feedback, express each generation's model, explanation, and validation quantity in that frame, and add two stylized models, one showing why mean aggregation dilutes the neighborhood signal of camouflaged fraudsters and one showing how verifier independence bounds the residual errors in a generated narrative. We find that statistical models and non-generative machine learning have mature validation practice anchored in supervisory guidance, that generative AI now appears both as a tool of fraudsters and as a component of defenses, and that in hybrid pipelines the generative component can write what an analyst or examiner reads although no dedicated supervisory framework covers it. We therefore formalize the evidence-traceability controls that can supply this missing validation as a division of verification labor, between the drafting agent, a deterministic field check, an independent verifier and the human examiner, governed by three parameters an institution can estimate; we then compare the generations quantity by quantity and close with a research agenda.
- Open Access
- 20 October 2026
- Published by MDPI AG in Preprints.org
Alert Triage via Drift-Adaptive Exploration for Human-in-the-Loop Anti-Money Laundering
Anti-money-laundering monitoring divides work between a model and a team of analysts: the model ranks transactions, analysts review as many alerts as their capacity allows, and the labels they return are the model’s only supervision. A budget spent on the highest-scoring alerts maximizes immediate seizures while leaving the model blind to patterns it has not yet learned. This exploitation-exploration trade-off is known in fraud detection, but how to adapt the split of a review budget to concept drift has received little attention. We propose DAT, a drift-adaptive triage policy that splits the review budget between a detection and an exploration channel. It selects the highest-risk transactions for detection, then transactions that are uncertain under the current model or distributionally novel for exploration, clustering the latter so the budget buys distinct information rather than near-duplicates, and adapts the split online from two drift signals: the analyst hit rate in the detection channel and the divergence between the current and the training distributions. Every alert carries the reason it was raised so that a team can tell a seizure case from a learning case, a design element we specify but do not evaluate, because our analysts are simulated. Replaying the labeled Elliptic Bitcoin transactions step by step with simulated analysts at a 10% budget and a 1-step label delay, DAT retains 52% of its pre-shift illicit-class F1 after the documented distribution shift, against 42% for a drift-triggered policy that widens exploration only once a detector fires and 36% for pure exploitation, while catching as many illicit transactions as either, 707 against 692 and 680, and at no measurable cost in pre-shift F1. The review budget is an enforcement resource and a learning channel at once, and DAT lets a team set how it is divided.
- 20 October 2026
- Published by MDPI AG in Preprints.org
Open AccessAlert Triage via Drift-Adaptive Exploration for Human-in-the-Loop Anti-Money Laundering
Anti-money-laundering monitoring divides work between a model and a team of analysts: the model ranks transactions, analysts review as many alerts as their capacity allows, and the labels they return are the model’s only supervision. A budget spent on the highest-scoring alerts maximizes immediate seizures while leaving the model blind to patterns it has not yet learned. This exploitation-exploration trade-off is known in fraud detection, but how to adapt the split of a review budget to concept drift has received little attention. We propose DAT, a drift-adaptive triage policy that splits the review budget between a detection and an exploration channel. It selects the highest-risk transactions for detection, then transactions that are uncertain under the current model or distributionally novel for exploration, clustering the latter so the budget buys distinct information rather than near-duplicates, and adapts the split online from two drift signals: the analyst hit rate in the detection channel and the divergence between the current and the training distributions. Every alert carries the reason it was raised so that a team can tell a seizure case from a learning case, a design element we specify but do not evaluate, because our analysts are simulated. Replaying the labeled Elliptic Bitcoin transactions step by step with simulated analysts at a 10% budget and a 1-step label delay, DAT retains 52% of its pre-shift illicit-class F1 after the documented distribution shift, against 42% for a drift-triggered policy that widens exploration only once a detector fires and 36% for pure exploitation, while catching as many illicit transactions as either, 707 against 692 and 680, and at no measurable cost in pre-shift F1. The review budget is an enforcement resource and a learning channel at once, and DAT lets a team set how it is divided.
- Open Access
- 10 October 2026
- Published by MDPI AG in Preprints.org
From Detection to Investigation: A Survey of Foundation Models and Agents for Financial Fraud and Financial Crime
Foundation models are extending research on financial fraud and financial crime from isolated risk detection to document understanding, relational analysis, and investigative workflows. A gap nevertheless separates model outputs from candidate evidence and actionable conclusions. This survey examines advances in six families of foundation models: language, vision, speech, temporal, graph, and tabular models. It compares their roles and limitations along two dimensions: the risk-control lifecycle and the directness of the supporting evidence. Existing studies show that pretraining can support representation reuse and the processing of unstructured information, with direct validation in some financial tasks. Cross-modal identity binding, information availability over time, report faithfulness, and case-level benefits require separate evaluation. Evidence on investigative agents spans prototypes, controlled tasks, expert assessments, and industry disclosures, which establish different aspects of workflow feasibility, task performance, and user assessment. Establishing gains in investigative efficiency and reductions in financial losses requires case-level comparative evidence. We examine entity and temporal alignment, the organization of candidate evidence, constrained tool execution, and investigative benchmark design, together with the implications of adversarial threats, privacy, fairness, and model risk governance. Research priorities should shift from expanding system functionality toward factual traceability, sufficient evidence for decisions, and reviewable actions.
- 10 October 2026
- Published by MDPI AG in Preprints.org
Open AccessFrom Detection to Investigation: A Survey of Foundation Models and Agents for Financial Fraud and Financial Crime
Foundation models are extending research on financial fraud and financial crime from isolated risk detection to document understanding, relational analysis, and investigative workflows. A gap nevertheless separates model outputs from candidate evidence and actionable conclusions. This survey examines advances in six families of foundation models: language, vision, speech, temporal, graph, and tabular models. It compares their roles and limitations along two dimensions: the risk-control lifecycle and the directness of the supporting evidence. Existing studies show that pretraining can support representation reuse and the processing of unstructured information, with direct validation in some financial tasks. Cross-modal identity binding, information availability over time, report faithfulness, and case-level benefits require separate evaluation. Evidence on investigative agents spans prototypes, controlled tasks, expert assessments, and industry disclosures, which establish different aspects of workflow feasibility, task performance, and user assessment. Establishing gains in investigative efficiency and reductions in financial losses requires case-level comparative evidence. We examine entity and temporal alignment, the organization of candidate evidence, constrained tool execution, and investigative benchmark design, together with the implications of adversarial threats, privacy, fairness, and model risk governance. Research priorities should shift from expanding system functionality toward factual traceability, sufficient evidence for decisions, and reviewable actions.