PaperlayerView exact revision on arXiv

Source-linked AI summary

Debunk and Infer: Multimodal Fake News Detection via Diffusion-Generated Evidence and LLM Reasoning

Kaiying Yan, Moyang Liu, Yukun Liu, Ruibo Fu, Zhengqi Wen, Jianhua Tao, Xuefei Liu

arXiv:2506.21557v1cs.CL

TL;DR

Fake news videos are difficult to assess, while LLM safeguards can restrict diverse misinformation-related generation and yield inconsistent downstream benefits. DIFND combines multimodally conditioned diffusion-generated debunking evidence with collaborative MLLM reasoning, and outperforms competing methods on FakeSV and FVC.

  • Problem

    Fake news videos complicate credibility assessment, and LLM safeguards can restrict diverse misinformation-related generation, producing minimal and inconsistent downstream benefits.

    Source: §I · pp. 1–2

  • Method

    DIFND generates multimodally conditioned debunking cues with diffusion and uses a multi-agent MLLM chain-of-debunk strategy for verification.

    Source: §I · p. 2, A. Overview · p. 3, C. Chain of Debunk · p. 5

  • Results

    DIFND outperforms competing neural network-based methods on the FakeSV and FVC datasets.

    Source: §I · p. 2, B. Main Results · p. 7

  • Takeaways & Limitations

    The framework integrates generated evidence and reasoning-rich assessments to support more accurate and trustworthy misinformation identification.

    Source: §I · p. 2

Abstract

from arXiv · show

The rapid spread of fake news across multimedia platforms presents serious challenges to information credibility. In this paper, we propose a Debunk-and-Infer framework for Fake News Detection(DIFND) that leverages debunking knowledge to enhance both the performance and interpretability of fake news detection. DIFND integrates the generative strength of conditional diffusion models with the collaborative reasoning capabilities of multimodal large language models (MLLMs). Specifically, debunk diffusion is employed to generate refuting or authenticating evidence based on the multimodal content of news videos, enriching the evaluation process with diverse yet semantically aligned synthetic samples. To improve inference, we propose a chain-of-debunk strategy where a multi-agent MLLM system produces logic-grounded, multimodal-aware reasoning content and final veracity judgment. By jointly modeling multimodal features, generative debunking cues, and reasoning-rich verification within a unified architecture, DIFND achieves notable improvements in detection accuracy. Extensive experiments on the FakeSV and FVC datasets show that DIFND not only outperforms existing approaches but also delivers trustworthy decisions.

I. INTRODUCTION

DIFND combines generated debunking evidence with collaborative MLLM inference to address the difficulty of detecting fake news videos and the scarcity of suitable debunking data. The introduction situates this approach amid multimodal detection and MLLM-based verification research.

A. Multimodal Fake News Detection

  • Fake news videos combine subtitles, visuals, speech, and background audio, making them harder to detect through unimodal analysis alone.Prior multimodal approaches extend from text–visual interactions to audio and video modalities.Source: §I · p. 1

I. INTRODUCTION

  • Debunking data is often scarce or absent, and existing resources are largely text- and image-based, leaving short-form video debunking underexplored.The introduction notes that manual fact-checks and counter-narratives limit the quantity and scope of available data.Source: §I · p. 1

A. Multimodal Fake News Detection

  • Prior MLLM-based detection work faces concerns about augmentation artifacts and bias, limited complex-context reasoning, and restricted or inconsistent generation benefits.The introduction also identifies a general lack of interpretability in model outputs.Source: §I · pp. 1–2

I. INTRODUCTION

  • DIFND combines conditional debunk diffusion, multi-agent chain-of-debunk reasoning, and feature fusion to support more accurate and trustworthy multimodal fake-news detection.Its diffusion model generates debunking representations from multimodal news content, while LLM-generated rationales and contextual cues augment the debunking dataset.Source: §I · p. 2

B. Diffusion Model

The related work covers diffusion-based text generation and augmentation, alongside LLM reasoning methods. DIFND’s overview combines multimodal video features, generated debunking cues, collaborative verification, and attentive fusion.

B. Diffusion Model

  • Diffusion research includes continuous latent-space text generation and synthetic-data methods for low-resource classification, missing modalities, and fairness.Examples include DiffuSeq[33], Diffusion-LM[35], DiffusionCLS[36], and methods for recovering missing modalities[37].Source: B. Diffusion Model · p. 2

C. Large Language Model

  • LLM reasoning approaches use intermediate reasoning steps, interleaved actions, or agent collaboration; fake-news methods add knowledge guidance, generated justifications, or supportive and contradictory reasoning.Examples include chain-of-thought prompting, ReAct, and LVAgent[41].Source: C. Large Language Model · pp. 2–3

A. Overview

  • DIFND’s overview routes multimodal video features through debunk diffusion and chain-of-debunk verification, then attentively fuses extracted or generated features for the final decision.The framework includes text, audio, keyframes, debunking text, comments, and publisher profiles; diffusion can generate cues when explicit debunking information is unavailable.Source: A. Overview · p. 3

B. Debunk Diffusion

Debunk Diffusion treats debunking as both refuting false claims and authenticating true information, then generates multimodally conditioned latent cues. Its compression network maps text features into a compact representation for diffusion.

B. Debunk Diffusion

  • Debunking includes both disproving misinformation and authenticating truthful information, supporting a more balanced dataset for evaluating news credibility.The paper contrasts this scope with the conventional focus on refuting false claims.Source: B. Debunk Diffusion · p. 4
  • The debunk diffusion network samples latent debunking cues conditioned on multimodal video information, including when an item has no explicit debunking data.Its four components are a compression network, conditional latent diffusion, multimodal conditioning, and refinement; generated features are used in subsequent fusion.Source: B. Debunk Diffusion · p. 4, Figure 2 · p. 4

1) Compression Network:

  • The compression network uses learned latent queries and cross-attention to reduce encoder-feature sequence length and dimensionality while preserving essential details.Mean-pooled latent representations feed a classifier trained with cross-entropy loss.Source: 1) Compression Network · p. 4

2) Conditional Latent Diffusion:

The conditional latent diffusion module learns to denoise compressed debunk-language features using fused multimodal video features as conditions. It samples from noise through successive denoising transitions to produce debunk latent representations.

2) Conditional Latent Diffusion:

  • The diffusion process maps noise toward data through state transitions; its schedule γ(t) decreases monotonically from 1 to 0.The process samples time t continuously over [0, T].Source: 2) Conditional Latent Diffusion · p. 4
  • The denoising network reconstructs compressed debunk-language data using a mean-squared-error loss and multimodal news-video features as conditions.The conditions are the fused text, audio, and visual features.Source: 2) Conditional Latent Diffusion · p. 4
  • Generation starts from a standard-normal latent variable and applies successive denoising transitions to produce a sample from the debunk latent space.The text describes transitions from zT through intermediate states to z0.Source: 2) Conditional Latent Diffusion · p. 4

3) Multimodal Conditioning Mechanism:

The framework conditions debunk-feature generation on multimodal content, refines denoised features, and uses LLM-generated debunk texts to broaden training examples.

3) Multimodal Conditioning Mechanism:

  • Learnable positional encodings and cross-attention embed concatenated text, audio, and visual features as conditions in each denoising-transformer layer.The conditioning representation uses the modality features [Ft, Fa, Fv].Source: 3) Multimodal Conditioning Mechanism · pp. 4–5

4) Self-Refinement Network:

  • The self-refinement network transforms denoised latent features into refined debunk features, trained with reconstruction and classifier losses.Mean-pooled denoised latent features are passed to a classifier to obtain the classifier loss.Source: 4) Self-Refinement Network · p. 5

5) Debunk Text Augmentation by LLM:

  • LLMs generate debunk texts from video text, related event news, and existing debunks, using five prompting styles and hybrid sampling.Ground-truth texts are prioritized when available, and generated texts supplement them to provide five debunk texts per training instance.Source: 5) Debunk Text Augmentation by LLM · p. 5

C. Chain of Debunk

Chain-of-debunk uses multimodal agents to analyze video content and produce a reasoned authenticity assessment; captioning supplies structured text, visual, and audio representations.

C. Chain of Debunk

  • The multi-agent MLLM framework extracts textual, visual, and audio information to produce LLM-V assessments aimed at improving detection accuracy and interpretability.Agents analyze on-screen text, keyframes, and audio separately before evaluating cross-modal consistency, source credibility, and factual accuracy.Source: C. Chain of Debunk · p. 5

1) Multimodal Data Understanding and Captioning:

  • Deepseek-R1 analyzes titles and on-screen text, while Qwen2.5-VL and GLM-4-Voice caption keyframes and audio, respectively.Event entities guide generation to keep outputs aligned with the subject matter and reduce semantic deviations.Source: 1) Multimodal Data Understanding and… · pp. 5–6
  • The captioning process combines text, vision, and audio inputs with modality-specific prompts to produce text analysis and visual and audio captions.The outputs are denoted It, Iv, and Ia, respectively.Source: 1) Multimodal Data Understanding and… · p. 6

2) LLM for News Verification:

The verification stage records judgments and reasoning from repeated multimodal analysis, and Veri-Verdict Fusion combines textual-debunk and multimodal branches for final prediction.

2) LLM for News Verification:

  • Each news sample undergoes three independent LLM veracity analyses, with judgments and reasoning encoded by BGE as chain-of-debunk features.Inputs include news text, text analysis, visual and audio captions, event entities, and a verification prompt.Source: 2) LLM for News Verification · p. 6

D. Veri-Verdict Fusion

  • Veri-Verdict Fusion combines textual-debunk and multimodal branches, using attention-based feature fusion within each branch and late fusion for the final judgment.Textual-debunk fusion gates original text, denoised debunk cues, and chain-of-debunk features; multimodal fusion cross-attends text with audio and vision.Source: D. Veri-Verdict Fusion · pp. 6–7
  • Cross-entropy losses for the final, multimodal, and textual-debunk predictions are combined as a weighted sum, with hyperparameters controlling component influence.The performance-comparison table covers DIFND and baselines on FakeSV and FVC; no cell values are supplied in these passages.Source: D. Veri-Verdict Fusion · p. 7, Table I · p. 7

IV. EXPERIMENTS

The experiments use FakeSV and FVC, two video datasets with fake, real, and debunked examples. DIFND is compared with multimodal detection methods and multimodal LLMs.

1) Dataset.:

  • FakeSV contains 1,827 fake, 1,827 real, and 1,884 debunked Chinese short videos with text, audio, images, comments, and user profiles.Source: 1) Dataset · p. 7
  • FVC contains 727 fake, 483 real, and 40 debunked multilingual videos, with textual news, publisher profiles, and comments.Source: 1) Dataset · p. 7

2) Baselines.:

  • The baselines include multimodal video detectors and GPT-4V, Qwen2.5-VL, and Qwen2.5-Omni.Source: 2) Baselines · p. 7

3) Implementation Details.:

Experiments compare DIFND with specialized detectors and zero-shot MLLMs, then examine implementation choices and component ablations. The reported results include cross-dataset gains and accuracy contributions from the framework’s modules.

3) Implementation Details.:

  • Feature extraction uses Qwen2.5-VL for visual captions, GLM-4-Voice for audio captions, Deepseek-R1 for text analysis, and BGE, XLM-RoBERTa, CLIP, and CLAP for features and alignment.Source: 3) Implementation Details · p. 7
  • FakeSV uses chronological 70:15:15 splits, while FVC uses five-fold cross-validation without event overlap; optimization uses Adam with a 0.0001 learning rate and batch size 64.Source: 3) Implementation Details · p. 7

B. Main Results

  • Zero-shot MLLMs perform below methods optimized for fake-video detection and show recall–precision imbalances associated with false alarms or missed detections.Source: B. Main Results · p. 7
  • In comparisons against neural network-based baselines, DIFND outperforms all competing methods, achieving 2.33% on the Fakesv dataset and 2.88% on the FVC dataset, respectively.Source: B. Main Results · p. 7

1) Effects of different components:

2) Cooperation of different modalities:

The paper reports that DIFND outperforms existing baselines on FakeSV and FVC. Its analyses find benefits from all three modalities, LLM-augmented debunk data, and balanced hyperparameter settings.

2) Cooperation of different modalities:

3) Effect of debunk data augmentation:

4) Effect of hyperparameters:

V. CONCLUSION

  • DIFND integrates generated debunking evidence and collaborative inference, and experiments on FakeSV and FVC report that it outperforms existing baselines.Source: §V · p. 8

APPENDIX

The appendix describes prompts for debunk-text augmentation and baseline MLLM detection, then illustrates detection cases from FakeSV.

1) Prompts of Debunking Text Augmentation for LLM:

2) Prompts of the Detection Task for baseline MLLMs:

B. Case Study

  • FakeSV cases illustrate how augmented debunk text adds evidence and reasoning connections, while chain-of-debunk checks internal and cross-modal inconsistencies.Source: B. Case Study · p. 12

Related papers

Loading 2506.21557v1…