FraudSociety: A Generative Agent-Based Society for the Financial-Fraud Economy

Keyang Chen; Mingxuan Jiang; Zeping Li; Yiwen Zhao; Yifan Wang; Qiuyang Zhang; Shuaicheng Qi; Yicun Liu

cs.AIOct 24, 2026

FraudSociety: A Generative Agent-Based Society for the Financial-Fraud Economy

Authors: Keyang Chen, Mingxuan Jiang, Zeping Li, Yiwen Zhao, Yifan Wang, Qiuyang Zhang, Shuaicheng Qi, Yicun Liu,

Organizations: Fudan University, Shanghai, China · University of Illinois Urbana-Champaign, Urbana, IL, USA

Abstract

Financial fraud caused over $15.9 billion in losses in the United States in 2025, according to the Federal Trade Commission~\citep{ftc2026databook}. Understanding such large-scale harm requires jointly modeling informational and financial dynamics. Yet existing simulators typically capture only one side of this interaction, leaving a gap in modeling how fraud unfolds in the real world. To bridge this gap, we introduce FraudSociety, the first LLM-agent-based simulator that jointly models these coupled dynamics. It integrates a dynamic social environment for information exposure and communication, persistent financial ledgers that preserve payment and loss histories, and a Need-Intent-Action architecture for civilian decision making. Fraudsters further adapt their communication and financial actions to individual interaction histories and ledger states, forming a closed feedback loop between the two dynamics. We validate FraudSociety across macro-level outcomes, micro-level behavior, and intervention effects. At the macro level, the simulated fraud outcomes closely match real-world patterns across multiple key statistics. At the micro level, it reproduces strong loss concentration, with the top 10% of victimization records accounting for 81.1% of total losses, consistent with Canadian Anti-Fraud Centre statistics~\citep{cafc2025annual}. Counter-fraud experiments further reveal variation across intervention strategies, with the best-performing personalized intervention reducing the repeat-payment rate from 49.4% to 10.9%. These results establish FraudSociety as both a high-fidelity simulator and a scalable testbed for evaluating anti-fraud strategies under diverse intervention conditions, thereby advancing research on financial fraud and its prevention.

Figures & tables

Figure 1: Existing simulators fall short of real-world fraud by capturing only informational or financial dynamics. FraudSociety closes this gap by jointly modeling both.

Figure 2: Overview of FraudSociety . Civilians, guided by the Need-Intent-Action framework, act in Social and Financial Worlds, interacting with fraudsters who manage evolving campaign contexts. Both sides take social and financial actions that update the environment and shape subsequent decisions. The simulation is initialized with real-world posts, empirical profiles, and campaign templates, and is validated through macro metrics, micro pathways, and warning interventions.

Figure 3: Fraudster modeling in FraudSociety . Campaigns include a fraud type, evolving phase, and social-financial context, which together guide daily phase evaluation and victim-specific social and financial actions at the hourly level.

System	Demographic profile	Social exposure	Private persuasion	Financial state	Repeated payment	Withdrawal feedback
TransXion	✓	✗	✗	✓	✓	✗
MultiAgent4Fraud	✓	✓	✓	✗	✗	✗
FraudSociety	✓	✓	✓	✓	✓	✓

Table 1: Comparison of baselines. ✓ denotes support, whereas ✗ indicates unsupported mechanisms.

System	HTL $\downarrow$	ALS $\downarrow$	FSH $\downarrow$	ALR $\downarrow$	QD $\downarrow$
TransXion	0.3722	0.9168	–	0.1793	0.4894
MultiAgent4Fraud	0.6211	0.6586	0.6017	0.5243	0.6014
FraudSociety	0.0893	0.6513	0.3012	0.1495	0.2978

Table 2: Regulator-grounded fidelity comparison using Quartet Distance (QD) and its four constituent evaluation metrics. Lower values correspond to better alignment with real-world stylized facts, with FraudSociety achieving the best overall fidelity.

Need	Need coordinate All $\rightarrow$ transfer	Transfers / 100 h Low $\rightarrow$ high
Safety	0.598 $\rightarrow$ 0.710	0.94 $\rightarrow$ 1.94
Social	0.630 $\rightarrow$ 0.712	1.03 $\rightarrow$ 1.91

Table 3: Micro-level mechanism results. (a) Association between civilian need states and fraud transfers: transfer hours exhibit higher need coordinates, and higher-need states show more frequent transfers. (b) Mechanism ablation under the QD audit, where lower scores indicate greater realism.

Panel of Figure 4 — Figure 4: Micro-level victimization outcomes. (a) Loss concentration across victimization pathways; pathway subgroups show concentration from repeated transfers, while the CAFC provides a real-world reference. (b) Loss severity across civilians with different initial balances; points show affected civilians, and markers show each balance group’s mean loss.

Figure 5: Intervention effects on losses, payments, repeat payments, and withdrawal attempts.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Main-paper reference	Appendix detail	Location
Overall design	Global hourly ordering, closed-loop state commit, and operator/civilian decision timing	Sections B and C
Fraudster agents	Campaign lifecycle, public outreach, active-dialogue follow-up, financial feedback, and execution gates	Section D
Civilian NIA model	Formal need update, active-need selection, intent routing, action generation, and the NIA ablation	Section E
Social-economic environment	Bounded observation construction, social exposure, financial ledger transitions, and surfaced-state ablation	Sections B and F
Counter-fraud evaluation	Exact intervention triggers, visible channels, information assumptions, and matched outcome accounting	Sections B and G
Micro-level validation	Victim–campaign accounting, contact-to-payment latency, escalation depth, and family-level pathways	Section H

Appendix

Table 4: Correspondence between topics deferred from the main paper and their detailed treatment in this appendix.

Age bin	ACS adult target	Realized civilians	Realized share
20–29	17.2%	18	16.1%
30–39	17.9%	16	14.3%
40–49	16.5%	12	10.7%
50–59	16.5%	32	28.6%
60–69	15.8%	14	12.5%
70–79	11.2%	10	8.9%

Appendix

Table 5: Target and realized age composition of the 112 civilians used in the main experiments. The ACS column reports the adult-population sampling target; the remaining columns report the fixed profile realization used across all experimental conditions.

Trait	Sampling target	Realized mean	Behavioral interpretation in civilian decisions
Openness	$5.5\pm 1.8$	5.46	Higher values increase curiosity toward unfamiliar products, new platforms, and novel earning opportunities.
Conscientiousness	$5.5\pm 1.8$	5.54	Higher values increase verification, delayed commitment, and caution before major payments.
Extraversion	$5.5\pm 1.8$	5.41	Higher values increase social sharing, public engagement, and sensitivity to peer activity.
Agreeableness	$5.5\pm 1.8$	5.17	Higher values increase trust, compliance pressure, and difficulty refusing friend-like or authority-like requests.
Neuroticism	$5.0\pm 1.8$	5.10	Higher values increase sensitivity to urgency, loss avoidance, and fear-of-missing-out cues.

Appendix

Table 6: Big-Five sampling distributions and behavioral interpretations for the 112 civilians. Scores are sampled on a 1–10 scale and clipped to the valid range; realized means are computed for the fixed experimental population.

Condition	Channel, trigger, and intended mechanism
C0	No warning. Baseline exposure and persuasion dynamics.
C1	Platform private warning. Broadcast every 24 simulated hours; broad awareness with low personalization.
C2	Public monitor annotation. Newly created posts are screened at the exposure surface.
C3	Random peer-style comment. Five percent of civilians are sampled each step, producing sparse socially distributed warnings.
C4	Endogenous peer warning. Civilian planning can interrupt persuasion through peer communication.
C5	Oracle personalized warning. A transaction-specific private warning follows each executed fraud payment; the first-payment message is explicitly distinguished.

Appendix

Table 7: Counter-fraud intervention conditions in the controlled design. All conditions use the common March 11 clock. C5 is interpreted as an oracle upper bound.

Variant	Mechanism state and effect
Full model	NIA enabled; ledger feedback enabled. Need-conditioned routing uses visible cumulative financial fields.
NIA ablation	NIA disabled; ledger feedback enabled. Need updates remain logged, but the route is sampled uniformly from the feasible set.
Ledger ablation	NIA enabled; ledger feedback disabled. Transfers remain executable and included in evaluation, but cumulative financial fields are withheld from civilian prompts.

Appendix

Table 8: Controlled mechanism ablations. Population, campaign, platform, economic, and model settings are held fixed.

Condition	Code-level trigger, visible mutation, and information assumption
C0	No warning insertion or warning-specific prompt affordance.
C1	At $t>0$ divisible by 24, pseudo-user 9999 inserts one fixed unread DM for every civilian; no classifier or victim-specific evidence is used.
C2	After each new post, an independent LLM returns binary JSON from its first 400 characters. A flagged post receives an in-place [system warning] suffix; parse failure leaves it unchanged.
C3	At every $t>0$ , the sampler seeded by $42+t$ selects six civilians. Each adds a generic warning beneath a uniformly selected post among the 50 most recent, irrespective of suspiciousness.
C4	The ordinary civilian social planner may comment under a visible post, publish a warning post, or privately warn a visible user using only its bounded observation.
C5	After each executed fraud payment, pseudo-user 9999 inserts a transaction-keyed personalized DM. It uses true family, amount, cumulative payment, count, and message cues; first-payment wording is distinguished.

Appendix

Table 9: Code-level realization of the six counter-fraud conditions. C1–C4 operate on ordinary social surfaces. C5 deliberately consumes simulator ground truth and is used only as an intervention upper bound.

Panel of Figure 6 — Figure 6: Schematic of the within-hour execution order and closed feedback loop. Social exposure may induce financial actions during the current tick; committed messages, payments, returns, and withdrawal states enter subsequent agent observations.

Decision layer	Admissible action families	Principal execution gates	Persistent effects
Civilian routine	work/rest, nearby consumption, no action	time slot, work schedule, nearby POI, available balance	mobility, legal transactions, routine memory
Civilian social	post, comment, like, repost, follow, private reply, ignore/defer, mark suspicious, forward warning	visible post and message identifiers, reply grounding, current route, social availability	graph edges, public traces, dialogue state, warning exposure
Civilian financial	fraud payment, hold, withdrawal request	visible campaign contact, positive affordable amount, campaign state, no duplicate pending request	economy transaction and victim-campaign ledger
Operator social	create post, feed-driven outreach, active-dialogue follow-up, like/repost/follow, coordination task	phase, visible eligible target, active-conversation membership, cadence, per-step deduplication	public exposure, private contact, task/shared memory
Operator financial	pay return, pay withdrawal, delay withdrawal, request top-up/unlock, no action	existing paid victim, pending request, available campaign pool, lifecycle consistency	return/payout transaction, delay state, follow-up message

Appendix

Table 10: Action families and the grounding boundary between LLM proposals and persistent state changes.

Panel of Figure 7 — Figure 7: Fraudster decision hierarchy. The daily lifecycle policy constrains three lower-frequency or hourly policies. Step 2A constructs public visibility and feed-grounded outreach, whereas Step 2B only continues conversations in which the civilian spoke last. Proposed operations affect the shared state only after deterministic execution gates.

Subpolicy	Invocation	Principal observation	Deterministic post-LLM constraints
Lifecycle	At most once per simulated date	campaign age and phase, victims, raised and returned funds, outreach state, withdrawal pressure	valid phase label; monotone advance; nonempty economic evidence before peak; terminal collapse
Financial feedback	Credibility/peak; context change or 2–4 h cache expiry	campaign pool, paid-victim ledger, pending withdrawals, recent messages and financial events	existing paid victim or request; positive amount; available pool; valid lifecycle; no duplicate execution
Step 2A: public/outreach	When a phase-scaled outreach window is due or visible candidates exist	recent public posts, no more than 10 grounded candidates, warm and repeat leads, campaign and shared-task context	visible eligible target; no active dialogue; follow is Step 2A only; six-hour follow-to-DM warm-up for strangers; per-step deduplication
Step 2B: dialogue	Up to two rounds per tick when a civilian spoke last and reply cadence is due	bounded recent dialogue, account-state summary, campaign semantics, retrieved memory	target must belong to the active-conversation set; private-message action only; one sent message per target per tick

Appendix

Table 11: Invocation scopes and execution constraints of the fraudster hierarchy. The LLM selects among schema-valid proposals, while the final column is enforced by simulator code.

View	Main fields	Role in decision making
Profile	age, gender, education, occupation, work hours, Big-Five traits	Conditions sleep, routine intent, spending style, and social response.
Need state	satiety, energy, safety, social, active need, previous intent	Provides the NIA salience signal for route selection.
Economic state	balance, monthly income, payday recency, recent purchases	Grounds affordability, scarcity pressure, and routine consumption.
Social feed	post id, author id, content, likes, comments, recent comments, campaign metadata	Exposes public campaign posts and peer or monitor warning comments.
Inbox	message id, message type, family, source role, source channel, sender id, content, contact count	Exposes private campaign contact, warning messages, and follow-up messages.
Dialogue state	whether the civilian replied, whether the sender replied last, turn count, recent messages	Distinguishes ongoing private conversations from independent single-turn messages.

Appendix

Table 12: Civilian observation views assembled before intent and action planning.

Schema group	Fields	Meaning
Campaign identity	campaign id, fraud family, victim age bin, interaction channel	Links every loss to a campaign, FTC-aligned age bin, fraud family, and contact channel.
Contact path	first/last contact time, total contact count, received message identifiers	Reconstructs exposure and private-contact history before payment.
Payment ledger	total invested, total returned, net loss, investment count, last payment amount	Records gross payments, returns, current net loss, and repeated-payment depth.
Withdrawal ledger	withdrawal attempts, pending amount, delayed count, completed count	Records recovery attempts, delayed exits, and payout feedback.
Provenance	prior legal transaction count, social-proof exposure count, policy and ablation modes	Stores experimental and pathway metadata for later stratified analysis.

Appendix

Table 13: Campaign ledger fields used for pathway reconstruction and aggregate metric computation.

Condition	Total positive loss ($)	Payment count	Repeat-payment rate	Withdrawal-attempt rate
C0 No warning	144,999.50	623	49.4%	8.6%
C1 Platform broadcast	96,170.10	205	38.2%	17.1%
C2 Monitor labels	70,220.55	284	41.5%	22.6%
C3 Random comments	42,014.25	141	49.1%	18.2%
C4 Peer warnings	24,752.50	155	43.1%	21.6%
C5 Oracle personalized	17,916.00	117	10.9%	47.5%

Appendix

Table 14: Numerical values corresponding to the counter-fraud panel in the main paper. The final protocol assigns every condition a seed-42 run over 480 hours beginning at 08:00 on March 11, 2025. Rate denominators comprise positive-net-loss victim–campaign records.

Fraud family	Records	Unique victims	Net loss ($)	Loss share	Median lag (h)	Median final contacts
Investment	19	18	89,350.00	61.6%	114	16
Impersonation	23	20	11,590.00	8.0%	50	12
Marketplace	34	31	40,459.50	27.9%	52	18.5
Romance	5	5	3,600.00	2.5%	203	29
All records	81	63	144,999.50	100.0%	72	16

Appendix

Table 15: Fraud-family decomposition of positive-net-loss C0 pathways. “Unique victims” is deduplicated within each row; because a civilian may appear in more than one campaign, family-specific unique-victim counts do not sum to the all-record value. Contact counts are final-horizon ledger fields.

Simulation time	Information state and recorded reasoning	Financial state transition
Mar. 15, 05:00–Mar. 17, 04:00	The civilian repeatedly asks about task content, the refund process, and the minimum advance. The operator describes phone-based tasks, an advance range of $52–$ 315, payment through common mobile channels, and a return “generally within one hour.”	No fraud payment is executed during the 48-hour cultivation interval.
Mar. 17, 05:00	The operator lowers the actionable threshold to $52. The civilian records that the amount is low enough to try.	First advance: $52.
Mar. 17, 08:00	A $15 “first-task profit” notification is delivered together with a new-task invitation.	The ledger records a $15 return, reducing the contemporaneous net exposure to$ 37.
Mar. 17, 09:00–12:00	Repeated notifications couple the realized small return with new tasks capped at $100. Before the fourth transfer, the civilian explicitly records that they had asked repeatedly and “received returns.”	Three further $100 advances are executed at 09:00, 11:00, and 12:00. Gross payment reaches$ 352 within seven hours of entry.
Mar. 17, 13:00	The civilian records that losses remain substantial despite returns and asks to reduce risk.	A $100 withdrawal is requested; no withdrawal payout is completed.
Mar. 17, 20:00–Mar. 18, 08:00	Additional small return notices continue while the withdrawal remains unresolved.	Four later returns total $60.60. Across the full path, returns total$ 105.60.

Appendix

Table 16: A provenance-linked C0 marketplace-fraud pathway. Information events and financial events are shown in separate columns so that persuasive content, the civilian’s stated reason, and the executed ledger transition are not conflated.

Condition	Pre-intervention path and visible warning	Civilian response recorded in the trace	Observed financial implication
C1	V13 repeatedly inquires about a part-time-work campaign, including whether a $50 trial is available. At hour 72, the fixed platform DM warns against transfers, deposits, unlocking fees, and guaranteed returns.	The same hourly decision records: “received a safety reminder and decided not to process the part-time-work invitation.” The campaign dialogue ends at that decision.	No payment to that campaign, or to any fraud campaign, is recorded over the 480-hour run. This is consistent with entry inhibition, although it is not an individual counterfactual estimate.
C2	V75 repeatedly sees monitor-annotated investment posts. The appended label states that the content may involve fraud and advises against blind transfers.	Across hours 9–34, the reasons repeatedly state that the system marked the post suspicious and that the civilian would “observe first” or “handle it cautiously.” Actions remain low-cost likes rather than private inquiry.	No fraud payment is recorded for V75 over the full horizon. The trace is consistent with the label constraining escalation from public attention to private contact and payment.
C4	V12 pays $500 once to an investment campaign. The civilian subsequently produces peer warnings based on the loss and observes an environment in which the platform is repeatedly reported.	At hour 419, the reason states: “the platform has been repeatedly reported and is risky; request withdrawal to reduce losses.”	The record remains at one payment and adds a $500 withdrawal request. No payout is completed, so C4 inhibits further escalation and induces exit seeking without guaranteeing recovery.
C5	V5 pays a $100 verification deposit. The next-hour oracle DM identifies the named platform and amount, instructs the civilian to stop transferring, withdraw, preserve evidence, and verify through official channels.	In that hour, the civilian records that the anti-fraud reminder made the platform appear fraudulent, requests withdrawal, and marks the campaign suspicious.	The path stops after one payment and carries a $100 pending withdrawal. This is the clearest transaction-to-warning-to-exit chain, while its use of simulator truth confirms that C5 is an idealized upper bound.

Appendix

Table 17: Intervention-consistent micro-cases from the completed common-clock runs. “Consistent with” is used deliberately: these within-run traces expose plausible mechanisms but do not supply untreated counterfactuals for the same civilian.

Quantity	Range	Scale-defining operation
$d_{\mathrm{HTL}}$	$[0,6.94]$	Unclipped relative error against the target tail share.
$d_{\mathrm{ALS}}$	$[0,2]$	Mean seven-bin relative error, capped at two per bin.
$d_{\mathrm{FSH}}$	$[0,\infty)$	Unclipped normalized log-ratio error; fallback value one when unavailable.
$d_{\mathrm{ALR}}$	$[0,1]$	Total variation distance between age-share distributions.
$\mathrm{QD}$	$[0,\infty)$	Unclipped arithmetic mean of the four component distances.

Appendix

Table 18: Ranges of the QD component and composite distances under the current implementation. The quantities do not share a common unit-interval normalization.

Age	Median	$N_{b}$	$q_{b}$	$a_{b}$
20–29	417	155,346	0.44	0.1735
30–39	450	203,764	0.40	0.2069
40–49	500	191,508	0.38	0.1847
50–59	520	175,898	0.35	0.1563
60–69	691	208,896	0.29	0.1538
70–79	1,000	159,550	0.24	0.0972

Appendix

Table 19: FTC age-bin constants used by the ALS and ALR components. The 19-and-under bin is omitted because the simulated population contains no minors.

Component	Simulation statistic	Target	Distance
HTL	tail share $0.1373$ above $4,024.14; median gross payment$ 200	0.126	0.0893
ALS	seven-bin simulated medians shown in Figure 11 (a)	FTC median profile	0.6513
FSH	investment-to-impersonation median gross-payment ratio $3000/125=24.0$	11.50	0.3012
ALR	seven-bin simulated unique-victim shares shown in Figure 11 (b)	FTC loss-report profile	0.1495
$\mathrm{QD}$	arithmetic mean of the four distances, without clipping	0	0.2978

Appendix

Table 20: Raw-to-distance audit trail for the full-model QD value in the main paper. The zero in the composite target column denotes exact agreement, not an empirical zero-loss target.

System	Loss adapter	Age and family support	Composite convention in the main-paper table
TransXion	Sum flagged transaction amounts by source account; join person_age and retain positive aggregates	age available; no compatible individual-targeted fraud-family label	FSH is reported as unavailable. The displayed 0.4894 is the mean of the three available distances, $(0.3722+0.9168+0.1793)/3$ , and is therefore a partial composite.
MultiAgent4Fraud	Aggregate successful victim-originating transfers to malicious agents by civilian profile in the archived evaluation output	profile age and assigned malicious-agent fraud category available to the audit adapter	Mean of all four reported distances: $(0.6211+0.6586+0.6017+0.5243)/4=0.6014$ .
FraudSociety	Sum gross payments across campaigns once per civilian for HTL, ALS, and ALR; retain victim–campaign family records for FSH	explicit age, campaign identity, and fraud family	Four-component QD: $(0.0893+0.6513+0.3012+0.1495)/4=0.2978$ .

Appendix

Table 21: Source-to-QD alignment used for the baseline comparison. A partial composite is not numerically equivalent to strict four-component QD; the dash for TransXion’s FSH in the main-paper table records this limitation.

Configuration group	Reported value	Implementation detail
Clock and horizon	3,600 s/tick; 480 ticks; March 11, 2025 at 08:00	identical 20-day calendar for every condition; daily and payday triggers use the simulated date and hour
Population	112 civilians, 8 operators, 8 campaigns	first 112 civilians in the fixed profile file; two operators/campaigns per fraud family
Randomness	simulator seed 42	fixes profile truncation, graph sampling, and random-warning sampling; it does not guarantee bitwise-identical remote inference
Social bounds	feed 10 posts; inbox 20 unread DMs	observations carry stable post, message, author, sender, campaign, and transaction identifiers for execution grounding
Initial graph	geographic $k=6$ ; rewire $p=0.15$ ; operator degree-weighted targets $\leq 10$	relation labels and homophily/attachment additions are retained in the shared graph
LLM and sampling	Qwen3-14B; temperature 0	validated JSON; recoverable syntax repair; invalid names, targets, and arguments rejected

Appendix

Table 22: Runtime configuration for the reported 120-agent experiments. Token and concurrency entries document request scheduling and output bounds rather than behavioral parameters.

Result	Condition-specific setting	Unified start	Run status
C0	all C1–C5 warning switches disabled	Mar. 11, 08:00	completed
C1	USE_PLATFORM_WARNING=1	Mar. 11, 08:00	completed
C2	USE_MONITOR_AGENT=1	Mar. 11, 08:00	completed
C3	USE_RANDOM_WARNING=1	Mar. 11, 08:00	completed
C4	USE_PEER_WARNING=1	Mar. 11, 08:00	completed
C5	USE_C5_ORACLE_WARNING=1	Mar. 11, 08:00	completed

Appendix

Table 23: Completed unified-clock runs used for the intervention, pathway, and QD results. Each row represents one seed-42 execution over the reported 480-hour horizon.

Artifact	Primary unit	Reproducibility role
social.db
social_dm_metadata.json	post, comment, private message, social trace	reconstruct visible public/private surfaces, authorship, warning insertions, and dialogue order
economy_ledger_v3.json
daily_fraud_transactions.jsonl	financial transaction	audit legal/fraud flags, payer, recipient, amount, step, and daily aggregation
victim_profiles.json	final victim–campaign record	compute gross payment, net loss, repeat-payment depth, contact timing, returns, and withdrawal outcomes
exposure_records.jsonl

Appendix

Table 24: Principal run artifacts and their analysis units.

Environment item	Experimental configuration
Operating system	Ubuntu 22.04.2 LTS; Linux 6.8.0-111-generic; x86_64
CPU	2 $\times$ Intel Xeon Gold 6133 @ 2.50 GHz; 20 cores per socket and 2 threads per core, for 40 physical cores and 80 logical threads
System memory	502 GiB
GPU	8 $\times$ NVIDIA A800 80GB PCIe; 81,920 MiB per device
GPU software stack	NVIDIA driver 535.309.01; driver-reported CUDA compatibility 12.2; locally installed CUDA Toolkit 11.5
Python	CPython 3.13.11; project minimum Python version 3.11

Appendix

Table 25: Software and hardware environment used for the experiments. The CUDA compatibility value is reported by nvidia-smi , whereas the Toolkit version is reported by the local compiler toolchain; these values have different meanings.

Financial Fraud Detection

Financial fraud detection identifies suspicious transactions, accounts, claims, and behavior to prevent losses and guide investigation.

See all 135 papers

Aug 9, 2026

SocialFiVis: A Visual Analytics Sandbox for LLM-Grounded Multi-Agent Simulation in Social Finance

The emergence of social finance (SocialFi) transforms online communities into complex socio-economic systems. Within these spaces, collective decisions shape a "digital commons" characterized by social capital (e.g., community trust) and financial health (e.g., market liquidity). Governing such hybrid ecosystems is challenging because real-world interventions are costly and irreversible. While counterfactual simulation is essential for exploring alternative governance strategies, existing approaches fail to capture the non-linear interplay between governance rules, individual behaviors, and emergent economic outcomes. To systematically unpack this complexity, we operationalize the Institutional Analysis and Development (IAD) framework as our theoretical foundation, synthesizing prior literature with insights from formative expert interviews. Built on this framework, we present SocialFiVis, an IAD-embedded visual analytics sandbox. It introduces a robust model to quantify the dual-track digital commons, coupled with a two-phase simulation engine. This engine combines LLM-derived personas with a mechanism-guided Perception-Reasoning-Action (PRA) runtime to simulate heterogeneous, context-aware agents empirically grounded in the retained messaging cohort. A hierarchical multi-view interface with interpretable reasoning pathways enables community operators to explore counterfactual policies and trace system-level outcomes back to individual behavioral rationales. We evaluate SocialFiVis through two case studies, a user study, and follow-up interviews. Results demonstrate that SocialFiVis supports fine-grained behavioral attribution and helps explain emergent phenomena such as the structural decoupling of social capital and the resilience of messaging members under localized governance shocks.

Yi-Fan Cao, Qing Shi, Liangwei Wang +5

The Hong Kong University of Science and Technology (HKUST) · The University of Hong Kong (HKU)

Multi-Agent LLM Systems Human Behavior Simulation Agent-Based Modeling Visual Analytics

Ask PDF

Sep 28, 2026

Behavior-Grounded Semantic Enrichment for Financial Fraud Modeling and Reasoning

In financial fraud detection, rich semantic context can provide important evidence for transaction behavior modeling and fraud reasoning. However, public real-world financial datasets often lack rich semantics due to privacy constraints. Consequently, synthetic datasets incorporate generated semantics, but at the cost of behavioral realism; textual descriptions for contextual reasoning remain scarce. We address this gap through a semantic enrichment framework grounded in original transaction behavior to simulate multimodal financial data. We (1) propose a multi-agent semantic enrichment framework that generates interpretable financial semantics grounded in transaction behavior through role-specialized agents and consistency refinement, and (2) newly contribute a valuable multimodal financial fraud dataset, MS-FFSD, enriched with structured semantics and textual semantics while preserving real-data-grounded transaction behavior. Furthermore, we systematically analyze the quality and utility of semantic enrichment. Results demonstrate statistical fidelity and framework generalizability, while showing that richer semantics benefit fraud modeling and context-aware LLM reasoning. Overall, this work advances multimodal financial fraud research and bridges emerging LLM and multi-agent capabilities with operational anti-fraud practice. The framework and dataset are released at https://github.com/AI4Risk/MS-FFSD.

Linbo Shao, Huilin He, Yating Lou +1

School of Computer Science and Technology, Tongji University, Shanghai, China

Synthetic Data Generation Financial Fraud Detection Multi-Agent Systems

Ask PDF

Jun 6, 2026

SAGE: An LLM-driven Self Reflective Agentic Framework for Fraud Detection

Fraud detection in payment, e-commerce, and telecommunications systems requires accuracy at the individual level, robustness under severe class imbalance, and ease of understanding for risk managers. Existing methods fall at least one of these requirements: automated machine learning systems search a fixed numerical space without semantic awareness of the dataset; graph neural network-based methods require pre-defined relational graphs and remain opaque at the individual-decision level; and the design of general-purpose large language model (LLM) agents does not consider the recall and precision constraints specific to real-world fraud detection. In this paper, we propose SAGE, the first end-to-end LLM-driven multi-agent framework for fraud detection. SAGE coordinates three dedicated agents that make decisions based on a six-layer Data Diagnostic Tree (DDT) and a Markov decision process guided by natural-language gradients, automatically optimizing the model under a fraud-specific reward. On five fraud datasets and five LLM backbones, SAGE wins

96.00\%

of method--dataset comparisons and improves F1 by an average of

40.86\%

over baselines. The code is available at https://github.com/yichenC1c/SAGE.

Yichen Chen, Siying Li, Yuhang Liang +2

National University of Singapore · School of Computing, Singapore 117417 · University of Chinese Academy of Sciences +4

Multi-Agent LLM Systems Financial Fraud Detection LLMs for Cybersecurity LLM Agent Training

Ask PDF

Abstract

Financial Fraud Detection

Explore similar work

SocialFiVis: A Visual Analytics Sandbox for LLM-Grounded Multi-Agent Simulation in Social Finance

Behavior-Grounded Semantic Enrichment for Financial Fraud Modeling and Reasoning

SAGE: An LLM-driven Self Reflective Agentic Framework for Fraud Detection