🎬CSI: An Investigative Multi-Agent Framework for Explainable Short Video Fake News Detection

Source code and dataset of the paper "CSI: An Investigative Multi-Agent Framework for Explainable Short Video Fake News Detection", which is accepted by Findings of the 64rd Annual Meeting of the Association for Computational Linguistics (ACL 2026).

---

Architecture Overview

🎯 Multimodal Forensics Unit

(Each agent simultaneously)

• Visual Analyst: Analyze each shot in the video, generating a structured visual analysis $P_{\text{vision}}$ from the shot type, source material, visual content, and subtitles.
• Acoustic Analyst: For the audio in each video, generate structured audio analysis results $P_{\text{audio}}$ based on the boundaries of the shots, covering four aspects: human voice type, background music, sound effects, and transcription.
• Intelligence Analyst: Summarize the structured titles $T_s$ that represent the core events in the titles and texts of short video news, and use the Google Search API to search for relevant official reports $E$ on the internet.

Output the case file (shoting script, structured titles and official reports)

🔄 Case Review Unit

• Review Team: Based on the case file, a review team consisting of three reasoning agents with different roles conducts discussions in three stages and outputs the final deliberation result $R$.
• Adjudicator:By comprehensively utilizing the case files and the review results, we can make the final decision on whether the original short video news content is true or false(Individual training).

🚀 Installation

1. Install required dependencies:

pip install -r requirements.txt

2. Set up environment variables:

export GOOGLE_API_KEY="your_google_api_key"
export OPENAI_API_KEY="your_openai_api_key"

3. Quick train Adjudicator:

# Train the Adjudicator using FakeSV
 python main.py  --dataset fakesv  --mode train

 # Train the Adjudicator using FakeTT
 python main.py  --dataset fakett  --mode train 

📊Dataset

We conduct experiments on two datasets: FakeSV and FakeTT.