Is Your LLM-Based Multi-Agent a Reliable Real-World Planner? Exploring Fraud Detection in Travel Planning

ACL ARR 2026 January Submission1383 Authors

Keywords: Travel Planning, Fraud Detection, LLM-based Multi-Agent

Abstract:

The rise of Large Language Model-based Multi-Agent Planning has enabled autonomous and collaborative task execution across diverse domains. However, these systems often rely on external platforms such as review sites and social media, which are highly vulnerable to fraudulent information including fake reviews or misleading descriptions. Such reliance introduces risks that may cause financial losses and undermine user trust.
In this work, we present **WandaPlan**, a plug-and-play fraud environment designed to expose and evaluate fraud vulnerabilities in multi-agent planning systems. Rather than offering a standalone agent framework, WandaPlan provides modular fraud-injection scenarios—including Misinformation Fraud, Team-Coordinated Multi-Person Fraud, and Level-Escalating Multi-Round Fraud—that can be easily integrated into existing agent pipelines with minimal modifications.
Through experiments, we demonstrate that current frameworks prioritize task efficiency at the expense of data authenticity, leaving them susceptible to manipulation. We further show that WandaPlan can be applied across different open-source planning frameworks, highlighting its generalizability as a reusable evaluation direction.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: Language Modeling; Resources and Evaluation

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 1383