Financial misinformation detection (FMD) tasks frequently confront the complexity of financial claims and complicated relationships in the facts, causing obstacles when Large Language Models (LLMs) attempt to clarify the truth of a claim, especially in offline scenarios due to the absence of the ability to check correctness by finding news on the internet. There have been many works dedicated to designing more appropriate prompts to enhance LLMs' performance in misinformation detection tasks manually, achieving reasonable results. In our work, we utilize a multi-agent framework to address the offline financial misinformation detection task by collecting linguistic feature analysis reports, comment analysis reports, and fact-checking reports related to the claim, and generating prompts with an iterative method according to historical performance and instances provided. After the reports and prompts are collected, we test the ability of the designed prompts with state-of-the-art LLMs, including GPT-3.5-turbo and Deepseek-V3.2-chat. The result demonstrates that the training method based on a multi-agent and iteratively generating prompts achieves a modest performance in the offline financial misinformation detection task.

PDF