Research project · 2026 Dataset gated · manual review

Can MLLMs generate—and detect—multimodal fake news?

A paired benchmark and multi-agent framework for realistic text–image misinformation—and the models asked to detect it.

Jiyao Yang · Yang Liu · Zhenyue Qin · Qingyu Chen · Xiuzhen Zhang*

Read on arXiv Dataset & code access
4,507 source events
3 news domains
16 MLLMs evaluated
2 prompting protocols
01 / Method

A collaborative generation loop.

Specialized agents preserve event anchors, construct a counter-narrative, create supporting visuals, and review the result.

00

Source event

Published reporting supplies the factual anchors. Original source material is not redistributed.

01

Story agent

Reframes a central claim while keeping the event recognizable.

02

Visual producer

Selects a visual strategy and synthesizes a matching image.

03

Critic agent

Checks plausibility, cross-modal fit, format, and visual quality.

iterate until approved
02 / Findings

Detection is not solved.
Reasoning alone is not enough.

The evaluation goes beyond a final true/fake label to examine the evidence models use—and where it diverges from human judgment.

01

Human–model gap

Most evaluated MLLMs remain substantially behind expert human judgments on multimodal news veracity.

02

Visual authenticity

Models may align images with text while still failing to determine whether the visual itself is authentic.

03

Uneven CoT effects

Structured reasoning improves many models but can also amplify model-specific bias and shift decision boundaries.

03 / Research access
Gated research access

One request.
Data and code.

Access is limited to non-commercial academic, educational, and scientific research, including reproduction and evaluation. Read the conditions below, then apply through Hugging Face. Each request is reviewed manually.

Temporary access noticeYou may submit an access request. Approvals are temporarily paused while legacy repository content is removed. Existing requests remain in the review queue.

Request access Read conditions
Individual, manual approvalApproval covers the released data and code for your account. Collaborators must apply separately.
Synthetic dataConstruction codeEvaluation code & prompts
Show BibTeX citation 
@inproceedings{yang2026multimodalfakenews,
  title  = {Can Multimodal Large Language Models Generate and Detect
            Multimodal Social Media Fake News?},
  author = {Yang, Jiyao and Liu, Yang and Qin, Zhenyue and
            Chen, Qingyu and Zhang, Xiuzhen},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year   = {2026},
  eprint = {2609.35809},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2609.35809}
}
How to apply

A short application.
A human review.

  1. Sign in to Hugging Face.Open the gated repository using your own account and select the access request option. Hugging Face shares your username and account email with the reviewers.
  2. Tell us who you are and your intended use.Provide your full name, university or institution, institutional email, a public university profile or ORCID link, and a brief description of your intended research use. A detailed research proposal is not required.
  3. Accept the conditions and submit.Hugging Face records your request in the review queue. Email notifications depend on the reviewers' notification settings and may be delayed; the queue is the authoritative record of your request. Reviewers may contact your institutional email to confirm your affiliation.
  4. Return to the same repository.Use the same Hugging Face account to check your request status and, once approved, download the released data and code. Submitting an application does not itself grant access.

Identity and affiliation are checked manually. Providing an email address or profile link does not by itself verify either.

Questions about an application?
xiuzhen.zhang@rmit.edu.au

Before requesting access

Conditions of use

  1. Non-commercial research only.The resources are limited to the approved academic, educational, or scientific research purpose. Commercial use is prohibited, including product development, commercial services, internal business operations, and monetization through third parties. Provide accurate identity, affiliation, and intended-use information; contact the project team before changing your research purpose.
  2. Keep access personal and controlled.Approval applies to the named applicant. Do not redistribute, sell, share, mirror, host, or transfer the code, data, derived unencrypted subsets, archives, passwords, credentials, tokens, or access links to third parties, public repositories, or external platforms. Each collaborator must obtain their own access through an individual request. Do not bypass the access controls.
  3. Do not use the resources to mislead or cause harm.Do not present synthetic content as authentic reporting or use it for deception, impersonation, targeted persuasion, harassment, fraud, election interference, or operational misinformation. Do not use the resources to target or harm individuals or groups.
  4. Protect the materials.Use secure storage and named-user access controls. Report suspected loss, disclosure, or unauthorized access promptly to xiuzhen.zhang@rmit.edu.au.
  5. Make research outputs traceable.Cite the paper and dataset/code revision. Clearly label synthetic or altered material, retain provenance and safety notices, and disclose material filtering, relabeling, preprocessing, and other changes.
  6. Publish findings responsibly.Aggregate results and limited excerpts necessary for scientific review may be published under the research license. Outputs must not substitute for the dataset, allow reconstruction of a substantial part of it, reveal access information, or facilitate harmful reuse.
  7. Respect source rights.Original news articles, original source images, and true-post texts are excluded from the release. Obtain any required third-party materials independently under their applicable terms. Access does not transfer rights held by publishers, authors, or other parties.
  8. Follow review and retention decisions.The team may request clarification, decline a request, or suspend or revoke access. Stop using and securely delete local, server, cloud, cache, and backup copies when the approved research period ends or access is revoked, subject to legally required retention.

Access is subject to the research use license and data use agreement. The materials are provided as-is and may contain errors or misleading content; approval does not endorse the applicant or their research.

Citation copied