LiarMP4 | Multimodal Content Moderation

Problem and Target Audience

Target User: Trust and Safety teams, investigative journalists, and social media platforms.

The Problem: Traditional content moderation relies on Predictive AI evaluating metadata such as account age and engagement velocity to output a scalar probability. While computationally efficient, this approach fails entirely against Contextual Malformation. This occurs when a completely authentic video is paired with a fabricated caption. Standard deepfake detectors pass the video as real, but the semantic intent is entirely deceptive.

We propose using Generative AI to extract distinct Veracity Vectors including Visual Integrity, Audio Integrity, Source Credibility, Logic, and Emotion to provide auditable and interpretable moderation signals.

Scope Boundaries

What We Built:

Fractal Chain of Thought Orchestrator: A recursive inference strategy to verify multiple modalities.
Human In The Loop Labeling Studio: A browser extension and UI to generate verified Ground Truth datasets.
TOON Parser: Strict Token Oriented Object Notation parsing for reliable structured data output.

What We Reused:

Foundation Models like Gemini 2.5 Flash, Gemini 3.0 Flash and Qwen3-VL for multimodal inference.
Google Agent Development Kit for agentic communication.
AutoGluon and XGBoost for our Predictive AI baselines.

Out of Scope: Real-time live stream moderation due to latency constraints and generation of synthetic deepfakes.

Key Results and Evaluation

Models were evaluated against our manually verified Ground Truth dataset. We calculate a Composite Mean Absolute Error measuring the absolute distance across eight multi-dimensional Veracity Vectors alongside overall Tag Accuracy.

Method	Iteration Depth	Composite Error	Tag Accuracy	Average Latency
Baseline Predictive AI	0	38.4	44%	1.2s
Baseline Generative AI	1	24.1	18.4%	15.4s
Generative AI Fractal Logic	2	12.8	64.8%	34.1s
Multi-Agent System	3	7.4	92.1%	45.8s

Empirical Model Leaderboard

Detailed performance metrics evaluated across our verified Ground Truth dataset.

Type	Model	Prompt	Reasoning	Tools	FCoT Depth	Accuracy	Comp. MAE	Tag Acc
GenAI	gemini-2.5-flash	standard	fcot	None	2	83%	15.11	64.8%
GenAI	gemini-2.5-flash	standard	fcot	Search, Code	2	77.3%	16.69	34.2%
GenAI	gemini-2.5-flash	standard	cot	Search, Code	1	76.2%	20.79	20.2%
GenAI	gemini-2.5-flash	standard	none	Search, Code	0	71.4%	23.33	22.2%
GenAI	gemini-2.5-flash	standard	cot	None	1	63.6%	26.16	32.7%
PredAI	Gradient Boost	Standard	none	None	0	63.3%	11.04	82%
GenAI	gemini-2.5-flash	standard	none	None	0	63%	20.66	18.4%
GenAI	gemini-2.5-flash-lite	standard	cot	None	1	56.5%	20.37	24.5%
GenAI	qwen3	standard	none	None	0	46.2%	44.36	27.8%
GenAI	gemini-2.5-flash-lite	standard	fcot	None	2	28.6%	27.21	30.7%
GenAI	qwen3	standard	cot	None	1	0%	54.44	80%

Detailed Vector Error Analysis

Breakdown of Mean Absolute Error across all eight veracity and alignment modalities.

Model	Prompt	Reasoning	Tools	Vis	Aud	Src	Log	Emo	V-A	V-C	A-C
gemini-2.5-flash	standard	fcot	None	6.38	12.13	17.02	11.28	14.47	29.15	10.85	23.4
gemini-2.5-flash	standard	fcot	Search, Code	5.91	13.18	16.36	15.00	16.36	30.91	12.27	22.73
gemini-2.5-flash	standard	cot	Search, Code	12.86	21.90	17.14	20.48	23.33	30.00	17.62	27.14

Key Takeaways & Future Directions

FCoT Efficacy: Deep, recursive reasoning like depth 2 without external tools achieves the highest overall accuracy (64.8%) and effectively maps videos into a stable latent space, producing a robust Tagging Solution that prevents rigid, misclassified pigeonholing.
External Tools: Counterintuitively, providing the agent with Web Search and Code Execution decreased accuracy; however, it increased Tag Accuracy (from 64.8% to 92.1%), as the agent had better context on text results but weighed raw visual/audio alignment vectors less.
Credibility Tracking: Expanding our User Credibility Profiler, future versions will persistently track "Account Honesty", maintaining historical integrity scores that penalize accounts repeatedly posting recontextualized media.
Video-Post Alignment Database: We are establishing a dedicated database to track specific video-post alignment patterns, allowing the system to cross-reference known "cheap fakes" and flag recycled authentic videos weaponized with deceptive captions.

Development Timeline and Iteration

Our approach evolved significantly as we identified the limitations of standard vision models against sophisticated real-world misinformation.

Phase 1: Open Source Vision

Our initial exploration utilized solely open-source Vision Language models such as Qwen3-VL. The primary goal was detecting spatial anomalies and deepfakes. We quickly discovered that high visual accuracy was insufficient for solving modern misinformation because genuine footage is routinely weaponized with false text.

Phase 2: Factuality Alignment

We integrated Dr. Ali Arsanjani's Factuality Factors (visit Alternus Vera) and Modality Alignment criteria. This shift allowed us to evaluate how audio, video, and text relate to one another, applying techniques like Veracity Vectors and Truthness Tensors. This successfully penalized malicious content where authentic videos were recontextualized with deceptive claims.

Phase 3: Multi-Agent System

To handle the complexity of processing isolated factuality vectors, we built a highly scalable architecture natively around the Google Agent Development Kit. This enables recursive verification steps and dynamic community context integration.

Alternus Vera & LiarMP4

Project Attribution: This work was developed as part of the Alternus Vera Research Project, focusing on LiarMP4. This project was conducted under the supervision of Dr. Ali Arsanjani.

alternusvera.com Core Codebase (DevKlim/LiarMP4)

Academic Citation

@misc{arsanjani_alternusvera,
  author = {Arsanjani, Ali and others},
  title = {Alternus Vera: A Research Project for LiarMP4, Detecting Contextual Malformation with Fractal Chain of Thought},
  year = {2024},
  publisher = {Alternus Vera Research Group},
  url = {https://alternusvera.com},
  note = {Core codebase: https://github.com/DevKlim/LiarMP4}
}