Probing Multimodal Large Language Models on Cognitive Biases in Chinese Short-Video Misinformation

2024-11-20 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

A new evaluation framework assesses Multimodal Large Language Models' (MLLMs) robustness against misinformation in Chinese short videos, using a manually annotated dataset of 200 videos from Douyin and Kuaishou. This dataset, covering four health domains, features fine-grained annotations for experimental errors, logical fallacies, and fabricated claims, verified by authoritative evidence. Eight frontier MLLMs, including Gemini-2.5-Pro, GPT-4o, and Qwen models, were evaluated across five input modalities. Gemini-2.5-Pro achieved the highest belief score of 71.5/100 in the multimodal setting, while o3 scored 35.2. The study revealed MLLMs are susceptible to cognitive biases: high video popularity surprisingly increased model confidence in their initial judgments, and authoritative channel IDs significantly influenced trust, often leading to lower performance on false content from verified sources. Multimodal inputs did not consistently outperform visual-only contexts.

Key takeaway

For AI Scientists and Machine Learning Engineers developing MLLMs for misinformation detection, recognize that current models are susceptible to cognitive biases like authority bias, where verified channels can deceptively influence trust. Your development efforts should focus on explicitly training models to be robust against such social cues and improving their ability to identify logical fallacies, as multimodal inputs alone do not guarantee superior performance. Prioritize bias mitigation and advanced reasoning over simply adding more modalities.

Key insights

MLLMs struggle with short-video misinformation due to cognitive biases like authority and popularity effects, despite multimodal reasoning capabilities.

Principles

• MLLMs exhibit authority bias, trusting verified channels regardless of content.
• High video popularity can reinforce MLLM's initial judgments, not necessarily induce false trust.
• Logical fallacies are particularly challenging for MLLMs to identify in videos.

Method

A framework evaluates MLLMs using a 200-video dataset with fine-grained error annotations (experimental errors, logical fallacies, fabricated claims) and a 7-level Likert scale for trustworthiness, calculating a normalized Belief Score (BS).

In practice

• Prioritize visual modality for short-video misinformation detection.
• Scrutinize MLLM outputs for biases related to source authority.
• Focus on improving MLLM's logical fallacy detection capabilities.

Topics

• Multimodal Large Language Models
• Misinformation Detection
• Cognitive Biases
• Short-Video Platforms
• Authority Bias
• Fact-Checking

Code references

• penguinnnnn/Fine-VDK

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Misinformation Span Detection in Videos via Audio Transcripts — Pinpointing misinformation within videos via audio transcript spans offers more granular, interpretable detection than video-level classification.
• A Multitask Transformer for Offensive Language Detection and Target Identification in HateBR — Multitask learning improves offensive language detection consistency but may reduce fine-grained target attribution.
• Cognitive Effects and Biases in Large Language Models — This tutorial clarifies cognitive effects and biases in LLMs by bridging psychology and NLP.
• Inside BullshitBench: AI Models and Nonsense Detection — AI models often confidently answer nonsensical questions, but specific training can significantly improve their ability to detect and push back against flawed premises.
• Week Ending 5.31.2026 — Multimodal LLM judges exhibit "Perceptual Judgment Bias," favoring plausible text over visual truth, which can be mitigated via targeted training.
• Scale Can't Overcome Pragmatics: The Impact of Reporting Bias on Vision-Language Reasoning — Reporting bias in VLM training data hinders reasoning, which scaling alone cannot resolve.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.