🤖 AI

CORE: Conflict-Oriented Reasoning for General Multimodal Manipulation Detection

The paper proposes CORE, a conflict-oriented reasoning framework that leverages a newly constructed Conflict Attribution Corpus to enhance multimodal large language models' ability to detect generalizable fake news by identifying intrinsic semantic or physical inconsistencies, thereby achieving superior performance in few-shot and zero-shot scenarios compared to existing state-of-the-art methods.

Original authors: Jinjie Shen, Yaxiong Wang, Yujiao Wu, Lechao Cheng, Tianrui Hui, Nan Pu, Zhihui Li, Zhun Zhong

Published 2026-06-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jinjie Shen, Yaxiong Wang, Yujiao Wu, Lechao Cheng, Tianrui Hui, Nan Pu, Zhihui Li, Zhun Zhong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the internet is a giant, bustling marketplace where people trade stories, pictures, and news. Lately, a new kind of merchant has arrived: Generative AI. These merchants are so skilled at painting pictures and writing stories that they can create "fake news" that looks and sounds completely real. It's like a master forger who can paint a portrait of a President winning a football trophy, even though that never happened.

The problem is that the old security guards (current detection tools) are trained to spot specific types of fakes. If they know how to spot a "crooked smile," they might catch that. But if the forger changes the trick to "weird lighting," the old guards are confused. They need to be retrained from scratch every time the forger changes their style, which takes too much time and data.

Enter CORE: The "Conflict Detective"

The authors of this paper, Jinjie Shen and his team, realized that all fake news has one thing in common: it doesn't make sense. It has an internal argument.

Think of it like a detective solving a mystery.

The Old Way: The detective memorizes the face of every known criminal. If a new criminal shows up with a different face, the detective fails.
The CORE Way: The detective learns the logic of a crime. They ask: "Does this story contradict what I know about the world?"

CORE (Conflict-Oriented REasoning) teaches a smart computer brain (called a Multimodal Large Language Model) to be that logic-based detective. Instead of memorizing specific fakes, it learns to spot conflicts.

How CORE Works: The Three-Step Training Camp

To teach this computer brain, the researchers built a special training program with three parts:

1. The "Conflict Library" (The Conflict Attribution Corpus)
First, they needed a textbook. They created a massive library of 14,000 fake news stories. But they didn't just label them "Fake." They broke them down like a forensic expert:

The Conflict Factor: What is the contradiction? (e.g., "The text says 'President' but the image shows a football award.")
The Conflict Source: Where did the lie come from? (e.g., "The text is lying," or "The image is lying," or "This contradicts real-world facts.")
Analogy: It's like giving a student a test where they don't just mark "Wrong," but write down exactly why the answer is wrong and which part of the question caused the error.

2. The "Bridge Builder" (Modality Bridging Pre-Training)
Computers often struggle to connect what they "see" (images) with what they "read" (text). Sometimes, the computer sees a picture of a football field but doesn't really "understand" the text saying "President."

CORE builds a bridge between these two worlds. It trains the computer to look at a picture and immediately understand the text description, and vice versa, so they speak the same language.

3. The "Logic Gym" (Conflict Perception Training)
Now, the computer goes to the gym. Using the "Conflict Library," it practices spotting contradictions.

The computer learns to push "conflicting ideas" far apart in its mind. If it sees "President" and "Football Award" together, it learns to scream, "Wait! These two things don't belong together!"
It learns to say: "This is fake because the text claims X, but the image shows Y, and my knowledge says X and Y can't happen at the same time."

Why This is a Game-Changer

The paper shows that this new approach is incredibly fast and flexible.

The "Few-Shot" Superpower: Imagine you want to teach a guard to spot a new type of fake news (e.g., deepfakes of a specific new celebrity).
- Old Methods: Need thousands of examples of that specific celebrity to learn.
- CORE: Only needs a handful of examples (sometimes just 100, or even zero!). Because it already understands the logic of conflict, it can instantly apply that logic to the new celebrity. It's like teaching someone to play chess by explaining the rules, rather than making them memorize every possible game.
The "Zero-Shot" Miracle: In some tests, CORE could spot brand-new types of fake news it had never seen before, just by reading the news and asking, "Does this make sense?"

The Bottom Line

The paper claims that by focusing on the intrinsic conflict (the logical contradiction) rather than the specific visual trick (the forgery technique), CORE creates a detector that is robust, general, and doesn't need a massive library of data to learn new tricks. It turns the computer into a critical thinker that asks, "Does this story hold up?" rather than just "Have I seen this picture before?"

The researchers have made their "Conflict Library" and their code available to the public, hoping others can use this "logic-based detective" to keep the internet honest.

Technical Summary: CORE – Conflict-Oriented Reasoning for General Multimodal Manipulation Detection

1. Problem Statement

The rapid advancement of generative AI has led to a surge in realistic and pervasive multimodal fake news, posing significant threats to public trust and social stability. Existing detection methods face two critical limitations:

Data Dependence: They rely heavily on large-scale, type-specific labeled datasets.
Design Rigidity: They utilize models and training paradigms tailored to specific manipulation types (e.g., face swapping, text-image inconsistency).

Consequently, these methods suffer from poor generalization when encountering emerging manipulation patterns. The "arms race" between forgery techniques and detection methods means that current models degrade significantly when faced with novel manipulation types, as the pace of new forgery evolution outstrips the cycle of data collection and model retraining. Furthermore, while Multimodal Large Language Models (MLLMs) possess rich world knowledge, they often lack the conceptual clarity to distinguish between semantically disparate concepts (e.g., conflating "U.S. President" with "Football Award" in feature space), leading to failures in robust misinformation detection.

2. Core Insight

The authors observe that the essence of manipulated misinformation lies in its intrinsic conflict. These conflicts manifest as:

Semantic Contradictions: Inconsistencies between content and common world knowledge (e.g., a news caption claiming a U.S. President won a football award).
Physical Inconsistencies: Discrepancies in lighting, shadows, or object properties between manipulated content and the original image or text.

The paper argues that if a detection model is endowed with explicit conflict-capturing capability, it can emulate human-like robustness, enabling effective adaptation to novel manipulation types with minimal data (few-shot or zero-shot) without relying on specialized designs for specific forgery traces.

3. Methodology: The CORE Framework

The Conflict-Oriented REasoning (CORE) framework is designed to equip MLLMs with explicit conceptual understanding and conflict detection capabilities. It consists of three main stages:

3.1. Conflict Attribution Corpus (CAC) Construction

To provide the necessary fine-grained supervision for conflict perception, the authors constructed the Conflict Attribution Corpus (CAC), containing 14,000 samples.

Source: Derived from the SAMM dataset (100k image-text pairs).
Annotation Process: Utilizing an expert pool of MLLMs (GPT-4o, Gemini2.5-Pro, Qwen3-VL-Plus), the system generates detailed rationales for why a news item is fake. These rationales are distilled into a structured format containing:
- Conflict Factors ( $C_1, C_2$ ): The specific contradictory concepts (e.g., "US President" vs. "Ballon d'Or").
- Conflict Sources ( $S_1, S_2$ ): The origin of the contradiction (Image, Caption, or World Knowledge).
Distribution: Conflicts originate from captions (29.98%), images (36.86%), and world knowledge (33.16%).

3.2. Modality Bridging Pre-Training (MBPT)

Since conflict annotations in CAC are textual but often describe visual elements, a Cross-Modal Aligner is trained to bridge the modality gap.

Data: 50k samples from the FineHARD dataset (image, positive text, and hard negative texts).
Mechanism: A cross-attention layer aligns visual features with text features.
Loss Function: Combines a contrastive learning loss (SigLIP) to ensure semantic correlation between matched text and visual features, and a language generation loss (VQA-style) to preserve language capabilities and fine-grained understanding.

3.3. Conflict Perception Training (CPT)

This stage explicitly reshapes the model's conceptual understanding to detect conflicts.

Input: Samples from CAC with structured conflict factors and sources.
Feature Processing: If a conflict factor originates from the image, the Cross-Modal Aligner extracts its visual feature; otherwise, the text feature is used.
Loss Functions:
- Conflict-Aware Contrastive Loss ( $L_{cacl}$ ): Pushes the representations of the two conflicting factors ( $z_1, z_2$ ) apart in the semantic space to establish clear conceptual boundaries.
- Conflict Reasoning Loss ( $L_{cr}$ ): A language modeling loss that trains the model to output the decision (Real/Fake) and the reasoning based on the conflict between $C_1$ and $C_2$ .

3.4. Rapid Adaptation

Once trained, the CORE framework allows for rapid adaptation to new manipulation types. The model requires only a simple question-answering instruction ("Is the news real or fake?") and minimal fine-tuning (few-shot) or zero-shot inference to achieve high performance on unseen manipulation patterns.

4. Key Contributions

General Paradigm: Introduction of a learning paradigm for general multimodal manipulation detection that rapidly adapts to novel manipulations with limited target samples, moving away from specialized designs for specific forgery types.
CORE Framework: A general framework that endows MLLMs with human-like conflict reasoning, enabling fast adaptation to unseen misinformation patterns by focusing on intrinsic conflicts rather than specific artifacts.
Conflict Attribution Corpus (CAC): A carefully curated dataset of 14k samples with fine-grained annotations of conflict factors and sources, serving as a benchmark for studying conflict reasoning in multimodal manipulation.

5. Experimental Results

The framework was evaluated using Qwen2.5VL-3B and Gemma3-4B as backbones across four datasets: DGM4, MDSM, MMFakeBench, and NewsCLIPpings.

Few-Shot Performance: With limited training data (100–750 samples), COREQwen and COREGemma achieved State-of-the-Art (SOTA) performance, surpassing the second-best methods by an average of 9.94% and 10.50%, respectively.
Zero-Shot Generalization: In cross-manipulation detection (training on one set of manipulation types and testing on unseen types), CORE significantly outperformed baselines (e.g., >14% improvement on MDSM).
Large-Scale Robustness: Even when trained on large-scale data (full SAMM, MDSM, and 200k NewsClippings), CORE maintained superior stability, outperforming baselines by ~3.8–4.2%.
Time-Sensitive Events: In evaluations involving 2025 events (outside the pre-training scope of the models), CORE achieved 74% accuracy, significantly outperforming baselines (44–52%), demonstrating its ability to detect logical violations without specific entity knowledge.
Ablation Studies: Experiments confirmed that both the MBPT stage (bridging modality gaps) and the CPT stage (learning conflict boundaries) are essential. Removing CPT caused a sharp performance drop (~4.58%), and removing the contrastive loss in MBPT led to a 4.44% drop.

6. Significance and Claims

The paper claims that CORE addresses the fundamental flaw of current detection approaches: their over-reliance on specific data patterns and model designs. By shifting the focus to the core conflict inherent in fake news, CORE enables models to achieve robust generalization.

Efficiency: The framework reduces the need for energy-intensive full-parameter retraining and large-scale, type-specific data collection.
Robustness: It provides a solution that remains effective against evolving generative AI threats, including zero-day misinformation where specific world knowledge is absent.
Ethical Considerations: The authors note that the CAC dataset is released strictly for research under a license prohibiting its use in generative tasks to prevent misuse in creating deceptive content.

The authors conclude that mastering the fundamental capability of conflict perception allows models to break free from data dependence, offering a scalable and robust approach to multimodal misinformation detection.