# REVEAL: Reasoning-enhanced Forensic Evidence Analysis for Explainable AI-generated Image Detection
**Authors**: Huangsen Cao, Qin Mei, Zhiheng Li, Yuxi Li, Ying Zhang, Chen Li, Zhimeng Zhang, Xin Ding, Yongwei Wang, Jing LYU and Fei Wu
> Huangsen Cao, Qin Mei, Zhiheng Li, Zhimeng Zhang, Yongwei Wang, Fei Wu are with Zhejiang University.
E-mail: huangsen_cao, yongwei.wang, wufei@zju.edu.cn.
Yuxi Li, Ying Zhang, Chen Li, Jing Lyu are with WeChat Vision, Tencent Inc.
Ding Xin is with Nanjing University of Information Science and Technology.
\NAT@set@cites
\justify
Abstract
With the rapid advancement of generative models, visually realistic AI-generated images have become increasingly difficult to distinguish from authentic ones, posing severe threats to social trust and information integrity. Consequently, there is an urgent need for efficient and truly explainable image forensic methods. Recent detection paradigms have shifted towards explainable forensics. However, state-of-the-art approaches primarily rely on post-hoc rationalizations or visual discrimination, lacking a verifiable chain of evidence. This reliance on surface-level pattern matching limits the generation of causally grounded explanations and often results in poor generalization. To bridge this critical gap, we introduce REVEAL-Bench, the first reasoning-enhanced multimodal benchmark for AI-generated image detection that is explicitly structured around a chain-of-evidence derived from multiple lightweight expert models, then records step-by-step reasoning traces and evidential justifications. Building upon this dataset, we propose REVEAL (R easoning- e nhanced Forensic E v id e nce A na l ysis), an effective and explainable forensic framework that integrates detection with a novel expert-grounded reinforcement learning. Our reward mechanism is specially tailored to jointly optimize detection accuracy, explanation fidelity, and logical coherence grounded in explicit forensic evidence, enabling REVEAL to produce fine-grained, interpretable, and verifiable reasoning chains alongside its detection outcomes. Extensive experimental results demonstrate that REVEAL significantly enhances detection accuracy, explanation fidelity, and robust cross-model generalization, benchmarking a new state of the art for explainable image forensics.
Index Terms: AI-generated image detection, Explainable AI, Forensic reasoning.
<details>
<summary>x1.png Details</summary>
### Visual Description
## Image Analysis: Image Authenticity Detection Diagram
### Overview
The image presents a diagram illustrating a three-stage process for determining whether an image is real or synthetic. It outlines the steps involved in analyzing an image from multiple perspectives, detecting evidence of authenticity or artificiality, and providing a final judgment.
### Components/Axes
The diagram is divided into three stages, each represented by a rounded rectangle:
* **Stage 1:** "Accept the user's instructions and analyze the image." This stage involves understanding the user's request to determine the authenticity of an image. It includes a speech bubble containing the user's request and an image of a puppy. A second speech bubble contains the statement "I understand the user's need. I will analyze and detect this image from eight different perspectives."
* **Stage 2:** "Performing multi-perspective, expert-informed image evidence analysis." This stage focuses on evidence detection. It lists four types of evidence:
1. Geometry flaws: The image was mistakenly classified as real.
2. Spectral clues: High-frequency artifacts were successfully detected.
3. High-pass fusion: The image was successfully detected as synthetic.
4. Local artifacts: The image was successfully detected as synthetic.
* **Stage 3:** "Provide an authenticity judgment based on the reasoning and analyze the findings across eight aspects." This stage involves reasoning and providing an answer based on the evidence gathered. It includes speech bubbles indicating the findings from each type of evidence, leading to a final answer.
### Detailed Analysis or ### Content Details
**Stage 1:**
* User's request: "Please help me determine whether this image is real or synthetic?...providing the reasoning conclusion."
* Image: A picture of a golden retriever puppy.
* Analyst's statement: "I understand the user's need. I will analyze and detect this image from eight different perspectives."
**Stage 2:**
* **Evidence Detection:**
1. **Geometry flaws:** "After geometric analysis, the image was mistakenly classified as real, as its plausible flow of fur, eye reflections..."
2. **Spectral clues:** "Through frequency analysis, the expert successfully detected high-frequency artifacts, with unexpected patterns in fur..."
3. **High-pass fusion:** "High-pass maps show that the expert successfully detected the image as synthetic, with inconsistent details such as overly sharp fur edges..."
4. **Local artifacts:** "The expert examined local pupil irregularities, successfully detecting the image as synthetic... Pixel-level anomalies..."
**Stage 3:**
* **Reasoning & Answer:**
* Spectral clues successfully detected high-frequency artifacts.
* High-pass fusion successfully detected the image as synthetic.
* Successfully detected anomalies based on shadow and lighting.
* The geometry flaws method mistakenly classified as real.
* Local artifacts successfully detecting the image as synthetic.
* Final Answer: `<answer>1</answer>`
### Key Observations
* The diagram outlines a structured approach to image authenticity detection.
* The process involves analyzing the image from multiple perspectives and gathering evidence to support a final judgment.
* The evidence detection stage focuses on identifying specific flaws and artifacts that may indicate whether the image is real or synthetic.
* The final answer is based on the reasoning and analysis of the evidence gathered.
### Interpretation
The diagram illustrates a multi-faceted approach to image authenticity detection, emphasizing the importance of analyzing an image from various perspectives and gathering evidence to support a final judgment. The process involves understanding the user's request, performing expert-informed image evidence analysis, and providing an authenticity judgment based on the reasoning and analysis of the findings. The diagram highlights the complexity of image authenticity detection and the need for a structured and comprehensive approach. The fact that most of the evidence points to the image being synthetic, while the geometry flaws were mistakenly classified as real, suggests that geometric analysis alone is not sufficient for determining image authenticity. The final answer `<answer>1</answer>` likely represents a binary classification, where 1 indicates that the image is synthetic.
</details>
Figure 1: Overview of the proposed REVEAL framework for reasoning-enhanced explainable synthetic image detection. The framework consists of three main stages: (1) receiving user instructions, (2) performing expert-grounded multi-perspective evidence detection, and (3) conducting reasoning through the chain of evidence (CoE) to derive a reliable decision with justifications.
1 Introduction
With the rapid evolution of generative artificial intelligence techniques such as Generative Adversarial Networks (GANs) [goodfellow2014generative, karras2019style] and Diffusion Models [dhariwal2021diffusion], the visual realism of synthesized content has advanced to a level that can easily deceive human perception. While these advanced models have unlocked unprecedented creative and economic potential in fields like digital art, design, and film production, they have also raised significant concerns regarding misinformation, privacy violations, and copyright issues. The continual progress in advanced diffusion models such as FLUX [black-forest-labs_flux_2024] and SD3.5 [esser2024scaling], along with autoregressive generation methods (e.g., VAR [tian2024visual]), has further intensified the challenge of distinguishing between real and synthetic content, making reliable detection an urgent research priority.
<details>
<summary>x2.png Details</summary>
### Visual Description
## Diagram: Explainable Classification vs. REVEAL
### Overview
The image presents two diagrams illustrating different approaches to classifying whether an image is real or synthetic. The first diagram, labeled "Explainable classification," outlines a straightforward process involving a Multimodal Large Language Model (MLLM) and thresholding. The second diagram, labeled "REVEAL," depicts a more complex process with multiple reward signals and group completion.
### Components/Axes
**Top Diagram (Explainable classification):**
* **Input:** An image of a bird on a branch, accompanied by the question "Please help me determine whether this image is real or synthetic?".
* **MLLM:** A green trapezoid labeled "MLLM" (Multimodal Large Language Model).
* **Thresholding:** A text box stating "Whether the model's prediction y > threshold τ".
* **Classification:** A yellow rectangle labeled "real/fake".
* **Explanation:** A yellow rectangle labeled "explanation" after "Next Token Prediction".
* **Background:** Light blue background.
**Bottom Diagram (REVEAL):**
* **Input:** An image of a bird on a branch, accompanied by the question "Please help me determine whether this image is real or synthetic?".
* **MLLM:** A green trapezoid labeled "MLLM" (Multimodal Large Language Model).
* **Outputs:** A column of outputs labeled o1, o2, ..., oG.
* **Reward Signals:** Three reward signals:
* R1: Answer reward (light blue)
* R2: Think Reward (light purple)
* R3: Multi-view alignment reward (light orange)
* **Group Completion:** A light blue rectangle labeled "Group Completion".
* **Evidence Analysis:** A yellow rectangle labeled "evidence analysis".
* **Classification:** A yellow rectangle labeled "real/fake".
* **Background:** Light yellow background.
### Detailed Analysis or ### Content Details
**Explainable classification:**
1. The process begins with an image and a question.
2. The image and question are fed into an MLLM.
3. The model's prediction (y) is compared to a threshold (τ).
4. Based on this comparison, the image is classified as "real" or "fake".
5. The model provides an explanation for its classification.
**REVEAL:**
1. The process begins with an image and a question.
2. The image and question are fed into an MLLM.
3. The MLLM generates multiple outputs (o1 to oG).
4. These outputs are evaluated based on three reward signals: Answer reward, Think Reward, and Multi-view alignment reward.
5. The outputs are used for group completion.
6. Evidence analysis is performed.
7. Based on the evidence analysis, the image is classified as "real" or "fake".
### Key Observations
* The "Explainable classification" approach is simpler and more direct, focusing on a single prediction and threshold.
* The "REVEAL" approach is more complex, incorporating multiple outputs, reward signals, and group completion.
* Both approaches aim to classify images as "real" or "fake" and involve an MLLM.
### Interpretation
The diagrams illustrate two different strategies for determining the authenticity of an image using multimodal large language models. The "Explainable classification" method represents a more traditional approach, where the model's prediction is directly compared to a threshold to make a classification, followed by an explanation. In contrast, the "REVEAL" method employs a more sophisticated approach by generating multiple outputs and using various reward signals to guide the model towards a more robust and accurate classification. The "REVEAL" method's use of group completion and evidence analysis suggests a more thorough and potentially more reliable classification process. The choice between these methods would likely depend on the specific application and the desired balance between simplicity and accuracy.
</details>
Figure 2: a) Existing post-hoc rationalization detection. b) REVEAL framework, a reasoning-enhanced paradigm for truly explainable forensic analysis.
Recent research [wang2020cnn, chai2020makes, wang2023dire, ojha2023towards, liu2024forgery, tan2024rethinking] has made notable progress in detecting AI-generated images. However, most traditional methods focus solely on discrimination, offering limited forensic analysis. The emergence of multimodal large language models (MLLMs) offers new opportunities, enabling models to combine visual perception with textual descriptions. Recent endeavours such as GPT-4 based detection [jia2024can], AIGI-Holmes [zhou2025aigi], FakeBench [li2025fakebench], and RAIDX [li2025raidx] have initiated this transition towards explainability. Yet, as illustrated in Figure 2, these methods share fundamental limitations: they primarily rely on post-hoc rationalizations or leverage the MLLMs merely as a powerful general-purpose visual classifier to identify high-level visual anomaly patterns (e.g.“unnatural lighting”, “blurry edge”). They fail to construct a causally grounded reasoning-based forensic pipeline where specialized evidence is systematically collected, analyzed, and synthesized through logical deduction. Specifically, these prior works: 1) use datasets (e.g. FakeBench [li2025fakebench]) that lack fine-grained, structured evidence, limiting support for deep causal reasoning; and 2) rely on methods (e.g. RAIDX [li2025raidx] with RAG) where explanations exhibit surface-level coherence derived from pattern matching, rather than being grounded in verifiable forensic evidence traces.
The critical gap highlights two major challenges in developing reasoning-enhanced synthetic image detection: 1) Lack of a reasoning-oriented forensic dataset. Existing datasets contain either binary labels or shallow textual justifications, without structured and rigorous chain-of-evidence annotations necessary to build auditable forensic judgments. 2) Limited reasoning-based explainability. Current MLLM-based detectors tend to produce post-hoc rationalizations instead of verifiable reasoning chains, leading to fragile generalization and unreliable claims in the forensic context.
To this end, we introduce REVEAL-Bench, a novel reasoning-oriented benchmark for AI-generated image forensics. Our data generation pipeline is fundamentally distinct from existing approaches: we shift from general visual correlation to expert-grounded evidence analysis. For each image, we first leverage eight lightweight expert models to provide structured, reliable, low-level forensic evidence. Such evidence then forms the input for a subsequent large model to generate a chain-of-evidence (CoE) annotation. By consolidating the multi-round forensic analysis from these specialized experts into a single, structured CoE trace, REVEL-Bench becomes the first dataset to explicitly provide an expert-grounded, verifiable forensic analysis that connects low-level cues to high-level conclusions.
Building upon this dataset, we propose the REVEAL framework, a two-stage training paradigm designed to enforce reasoning-based forensic evidence analysis. In the first stage, we employ a supervised fine-tuning (SFT) to teach the MLLM the canonical CoE structure. In the second stage, we introduce R-GRPO (Reasoning-enhanced Group Relative Preference Optimization), an expert-grounded policy optimization algorithm, featuring a novel reward function critical for enhancing the logical coherence and verifiability of forensic analysis. Specifically, R-GRPO jointly optimizes (i) detection accuracy, (ii) reasoning stability, and (iii) multi-view consistency. The novel optimization enforces the MLLM to perform logical synthesis over explicit forensic evidence rather than simple visual pattern matching, thereby achieving accurate, reliable, and explainable forensic analysis.
In summary, our work makes three major contributions:
REVEAL-Bench. We pioneer the first reasoning-based and explainable dataset for AI-generated image detection. Unlike prior datasets that offer only post-hoc explanations, REVEAL-Bench is uniquely structured around expert-grounded, verifiable forensic evidence that embeds an explicit chain-of-evidence following a systematic evidence-then-reasoning paradigm.
REVEAL Framework. We introduce the REVEAL Framework, a progressive two-stage training paradigm designed to instill standardized and explainable reasoning in multimodal LLMs. Its core, R-GRPO, optimizes the MLLM to perform logical synthesis for forensic evidence, jointly enhancing accuracy, reasoning consistency, and generalization.
Empirical Performance. Our approach achieves superior detection accuracy, generalization, and explanation fidelity, benchmarking a new state of the art for reasoning-based forensic research.
2 Related Work
Detection of AI-Generated Fake Images
The rapid evolution of generative models, e.g., GANs [goodfellow2014gan, esser2021taming], autoregressive models [oord2017vqvae], diffusion-based models [esser2024rectifiedflow, song2020ddim, ho2020ddpm, gu2022vqdiffusion, saharia2022imagen, ji2025mllm], has driven AI-generated images to near-photorealistic quality, challenging conventional detection methods. Early forensic studies focused on traditional manipulations like splicing or copy-move, analyzing noise inconsistencies, boundary anomalies, or compression artifacts [zhou2018manipulation, li2022splicing]. Researchers then shifted focus to generation artifacts, such as up-sampling grid effects, texture mismatches, or abnormal high-frequency decay [frank2020frequency, liu2020texture, dzanic2020fourier]. For example, the Spectral Learning Detector [karageorgiou2025spectral] models the spectral distribution of authentic images, treating AI-generated samples as out-of-distribution anomalies, achieving consistent detection across generators. However, as generators incorporate post-processing techniques like super-resolution, these low-level statistical clues become increasingly subtle and less reliable for robust detection.
Recent methods employ general-purpose feature extractors, such as CNN- or ViT-based detectors, to learn discriminative features directly. While lightweight CNNs achieve strong benchmark performance [ladevic2024cnn], methods like the Variational Information Bottleneck (VIB) network [zhang2025vib] aim to enhance generalization by constraining feature representations through the information bottleneck principle to retain only task-relevant information. Post-hoc Distribution Alignment (PDA) [wang2025pda] attempts to improve robustness to unseen generators by aligning regenerated and real distributions to detect unseen generators. Recently, NPR [tan2024rethinking] has become a representative approach by capturing low-level artifacts, demonstrating strong generalization capability. Similarly, HyperDet [cao2024hyperdet] and AIDE [yan2024sanity] achieve robust generalization through high-frequency spectrum analysis. Despite their discriminatory power, these approaches remain limited in forensic value, as their conclusions rely on global statistics and lack the semantic, verifiable evidence required for comprehensive explainability.
Explainable AI-generated Image Detection
The emergence of MLLMs [liu2023visual, wang2024qwen2] has accelerated the development of explainable image forensics by leveraging their advanced cross-modal understanding [wu2024comprehensive, talmor2019commonsenseqa]. Early efforts reformulated detection as a Visual Question Answering (VQA) task [jia2024can, keita2025bi, chang2023antifakeprompt], allowing MLLMs to provide accompanying descriptive text. FatFormer [liu2024forgery] extended this with a forgery-aware adapter to improve generalization on the CLIP-ViT [radford2021learning] encoder.
Subsequent studies focused on constructing task-specific multimodal datasets for fine-tuning. FakeBench [li2025fakebench] and LOKI [ye2024loki] provide synthetic images with manually written, high-level forgery descriptions. Holmes-Set [zhou2025aigi] utilized small models for initial image filtering and a Multi-Expert Jury mechanism to generate postt-hoc explanatory texts. At the methodological level, FakeShield [xu2024fakeshield], ForgerySleuth [sun2024forgerysleuth], ForgeryGPT [liu2024forgerygpt] and SIDA [huang2025sida] fine-tune MLLMs to achieve explainable forgery detection and localization. AIGI-Holmes [zhou2025aigi] integrates low-level visual experts with reasoning modules. RAIDX [li2025raidx] combines retrieval-augmented generation (RAG) [lewis2020retrieval] with GRPO optimization to improve the ability to describe texts.
Critically, existing datasets and methods suffer from two key limitations: First, the explanations are attributed to post-hoc rationalizations, often relying on the MLLM’s general knowledge and visual classification capabilities, failing to achieve logical synthesis of specialized forensic evidence. Second, they lack structured, fine-grained forensic evidence required to support a verifiable causal link between low-level artifacts and the final forensic judgments.
3 REVEAL-Bench
<details>
<summary>x3.png Details</summary>
### Visual Description
## AI-Generated Image Detection Workflow
### Overview
The image presents a workflow for detecting AI-generated images. It outlines the steps from data curation and pre-filtering to expert-grounded evidence collection, chain-of-evidence synthesis, and a final determination of whether an image is synthetic. The workflow incorporates both automated and expert analysis techniques.
### Components/Axes
**1. Data Curation & Pre-filtering (Top-Left)**
* **Title:** Data Curation & Pre-filtering
* **Sub-categories:**
* Chameleon Fake2M
* Autoregressive GAN
* GenImage
* Diffusion
* **Arrow:** A downward arrow labeled "Pre-filtering" points to the next stage.
**2. Expert Filtering (Lightweight Model as Expert) (Left)**
* **Title:** Expert Filtering (Lightweight Model as Expert)
* **Categories:**
* Local artifacts
* Spectral clues
* Pixel noise
* Spatial consistency
* Geometry flaws
* Shadow logic
* Texture fusion
* High-pass fusion
**3. Image Dataset Distribution (Bottom-Left)**
* **Type:** Pie Chart
* **Categories:** Real, Fake
* **Values:**
* Real: 30000
* Fake: 30000
* **Sub-categories (around the pie chart):**
* text: 6537
* scientific: 6
* scene: 680
* remote: 0
* object: 1158
* medical: 186
* image_text: 4553
* abstract: 4060
* animal: 52
* architecture: 97
* art: 456
* face: 4
* human: 482
* hybrid: 10782
* human: 77
* face: 4
* art: 22
* architecture: 22
* animal: 4724
* abstract: 588
* image_text: 6250
* medical: 6059
* object: 177
* remote: 1182
* scene: 1
* scientific: 6859
* text: 6537
**4. Expert-grounded Evidence Collection (Center)**
* **Title:** Expert-grounded Evidence Collection
* **Categories:**
* Local artifacts
* Spectral clues
* Pixel noise
* Spatial consistency
* Geometry flaws
* Shadow logic
* Texture fusion
* High-pass fusion
* **Prompt Design:** An icon of a person with a pen and paper is labeled "prompt design."
* **Arrows:** Arrows connect the categories to the "prompt design" icon.
**5. LLM Analysis (Top-Right)**
* **Statements:**
* "The expert successfully detected that the image is synthetic. Please analyze the local artifacts in the image." (Green checkmark)
* "The expert successfully detected that the image is synthetic. Please analyze the forgery using spectral clues." (Green checkmark)
* "The expert failed to detect that the image is synthetic. Please analyze its authenticity using high-pass fusion." (Red X)
* **LLM Icon:** A brain-like icon labeled "LLM" (Large Language Model) is present.
* **Explanations (connected to the LLM icon):**
* "Local artifacts: By observing the bird's eyes, we find that the reflection of the eyeball is missing..."
* "Spectral clues: Periodic artifacts of the synthesized image are revealed along the spectral axis..."
* "High-pass fusion: By examining the high-frequency map, it is observed that the area around the bird appears smooth and contains no signs of forgery..."
**6. Chain-of-Evidence Synthesis (Bottom-Center)**
* **Title:** Chain-of-Evidence Synthesis
* **Label:** visual evidence
* **Content:** Four image examples are shown.
**7. Reasoning Process (Bottom-Right)**
* **Title:** `<think>`
* **Steps:**
1. "Initial observation checks texture and lighting anomalies." (Magnifying glass icon)
2. "Detailed inspection identifies uniform surfaces and missing imperfections." (Pencil icon)
3. "Spatial analysis compares object-background alignment and projection logic." (Triangle icon)
4. "Shadow consistency test detects overly perfect lighting patterns." (Lightbulb icon)
5. "High-frequency analysis examines fine-grain texture irregularities." (X icon)
6. "Frequency spectrum evaluation reveals abnormal energy distributions." (Bar graph icon)
7. "Synthesizing all clues, the image is determined to be synthetic." (Checkmark icon)
* **Answer:** `<answer>1</answer>`
**8. Consolidation (Right)**
* **Arrow:** A large blue arrow points downward, labeled "consolidate," indicating the final step of combining all evidence.
### Detailed Analysis or ### Content Details
* **Data Curation & Pre-filtering:** This stage involves selecting and preparing the image data for analysis, using techniques like Chameleon Fake2M, Autoregressive GAN, GenImage, and Diffusion.
* **Expert Filtering:** This stage uses a lightweight model to filter images based on various features, including local artifacts, spectral clues, pixel noise, spatial consistency, geometry flaws, shadow logic, texture fusion, and high-pass fusion.
* **Image Dataset Distribution:** The pie chart shows a 50/50 split between real and fake images in the dataset, with each category containing 30,000 images. The surrounding labels indicate the distribution of different types of images within the dataset (e.g., text, scientific, scene, remote, object, medical, image_text, abstract, animal, architecture, art, face, human, hybrid).
* **Expert-grounded Evidence Collection:** This stage involves collecting evidence based on the categories identified in the expert filtering stage. The "prompt design" icon suggests that prompts are designed to elicit specific information from the images.
* **LLM Analysis:** The LLM analyzes the evidence and provides explanations for its conclusions. The examples show that the LLM can successfully detect synthetic images based on local artifacts and spectral clues, but may fail when using high-pass fusion.
* **Chain-of-Evidence Synthesis:** This stage involves combining all the evidence collected to make a final determination of whether an image is synthetic.
* **Reasoning Process:** This stage outlines the steps involved in the reasoning process, from initial observation to synthesizing all clues. The final answer indicates that the image is determined to be synthetic.
### Key Observations
* The workflow combines automated and expert analysis techniques.
* The LLM is able to successfully detect synthetic images based on certain features, but may fail when using other features.
* The chain-of-evidence synthesis stage is crucial for combining all the evidence and making a final determination.
### Interpretation
The workflow demonstrates a comprehensive approach to detecting AI-generated images. By combining data curation, expert filtering, LLM analysis, and chain-of-evidence synthesis, the workflow aims to provide a reliable method for identifying synthetic images. The use of both automated and expert analysis techniques allows for a more thorough and accurate assessment of image authenticity. The workflow highlights the importance of considering multiple factors and combining different types of evidence when detecting AI-generated images. The LLM's ability to provide explanations for its conclusions adds transparency and interpretability to the process. The workflow suggests that detecting AI-generated images is a complex task that requires a multi-faceted approach.
</details>
Figure 3: The pipeline of REVEAL-Bench. This figure illustrates our data processing pipeline, which consists of three stages: Data Curation & Pre-filtering, Expert-grounded Evidence Collection, and Chain-of-Evidence (CoE) Synthesis
As illustrated in Figure 3, this study constructs the REVEAL-Bench dataset through a rigorous, three-stage pipeline designed for reasoning-based image forensic: Data Curation & Pre-filtering, Expert-grounded Evidence Collection, and Chain-of-Evidence (CoE) Synthesis. This approach is fundamentally distinct as it replaces manual, subjective labeling with a process that systematically integrates verifiable evidence from specialized models with the logical synthesis capabilities of large vision-language models. The resulting dataset contains explicit, expert knowledge-grounded Chain-of-Evidence annotations, which is crucial for training forensic detectors with superior transparency and generalization capability.
Data Curation & Prefiltering
To ensure sufficient content, generator, and artifact diversity, we aggregate several prominent AI-generated detection benchmarks, including CNNDetection [wang2020cnn], UnivFD [ojha2023towards], AIGCDetectBenchmark [zhong2023patchcraft], GenImage [zhu2023genimage], Fake2M [lu2023seeing], and Chameleon [yan2024sanity]. This yielded in an initial corpus of approximately 5,120K synthetic images and 850K authentic images. To manage annotation costs while ensuring high data quality, we implemented a stratified sampling strategy based on automated quality assessments [talebi2018nima] and image resolution. Specifically, we sampled images based on aesthetic scores (50% high, 30% medium, 20% low), and image resolution, high-resolution ( $≥$ 512 $×$ 512) images at 50%, medium-resolution (384 $×$ 384–512 $×$ 512) images at 30%, and low-resolution ( $<$ 384 $×$ 384) images at 20%. Images were also semantically classified into 13 major categories (e.g., humans, architecture, artworks). After rigorous multi-stage filtering and preprocessing to eliminate non-representative or low-quality samples, we finalized a balanced corpus of 30K synthetic and 30K real images, which serves as the foundation for subsequent expert annotation
Expert-grounded Evidence Collection
To enable fine-grained, verifiable forensic analysis, we design and employ a set of eight lightweight and specialized expert models [li2025improving, sarkar2024shadows, tan2024rethinking, cao2024hyperdet, tan2024frequency, li2025optimized], each dedicated to screening and localizing a distinct category of synthetic artifact (as depicted in Figure 3). This is a crucial distinction from prior work, such as AIGI-Holmes [zhou2025aigi], which uses experts primarily for global filtering. Our experts, by contrast, provide structured, machine-readable evidence, including artifact masks and diagnostic labels. These eight outputs constitute the necessary forensic evidence foundation. By conditioning the LVLM on these high-fidelity, structured references, we ensure the final generated explanations are faithful, logically consistent, and verifiable against objective, low-level artifact data. This expert-grounded decompositional analysis effectively bridges the gap between small-model perception of artifacts and large-model logical reasoning.
<details>
<summary>x4.png Details</summary>
### Visual Description
## Diagram: CoE Tuning and R-GRPO Stages
### Overview
The image illustrates a two-stage process involving CoE (Chain of Evidence) Tuning and R-GRPO (Reward-Guided Policy Optimization). The diagram outlines the flow of information and the reward mechanisms used in each stage to determine whether an image is real or synthetic.
### Components/Axes
**Stage 1: CoE Tuning (Left Side)**
* **Title:** Stage 1: CoE Tuning
* **Input:** Two images (a bird on a branch and a prompt "Please help me determine whether this image is real or synthetic?"), each associated with a set of colored squares (orange and yellow).
* **Process:** The images and prompt are fed into an MLLM (Multi-modal Large Language Model).
* **Output:** The MLLM generates a text output containing both "think" and "answer" components.
* **Loss Functions:** L_think and L_answer are associated with the "think" and "answer" components, respectively.
**Stage 2: R-GRPO (Right Side)**
* **Title:** Stage 2: R-GRPO
* **Input:** An image (a bird on a branch) combined with the prompt "Please help me determine whether this image is real or synthetic?".
* **Process:** The combined input is fed into an MLLM, which generates multiple completions (Completion 1, Completion 2, ..., Completion G).
* **Reward Mechanisms:**
* **(1) Answer Reward:** Compares the "answer" component of a completion to a ground truth.
* Match: R = 1
* Mismatch: R = 0
* **(2) Think Reward:** Evaluates the "think" component of a completion.
* Match: R = 1
* Similar: R = 0.5
* Mismatch: R = 0
* **(3) Multi-view Alignment Reward:** Compares the "think" component with visual evidence.
* Match (example: structural irregularities and high-frequency artifacts): R = 1
* Mismatch (example: natural appearance but high-pass artifacts): R = 0
### Detailed Analysis or ### Content Details
**Stage 1: CoE Tuning**
* The input images are accompanied by colored squares. The left image has four orange squares, and the right image has four yellow squares.
* The MLLM generates a text output:
* `<think>1. Initial observation suggests...2. Upon closer inspection, uneven features are detected, indicating possible synthetic traces...7. After synthesizing all cues, the image is determined to be synthetic</think>`
* `<answer>1</answer>`
**Stage 2: R-GRPO**
* The MLLM generates multiple completions.
* **(1) Answer Reward:** The completion's answer is compared to a ground truth. A match results in a reward of 1, while a mismatch results in a reward of 0.
* **(2) Think Reward:** The completion's reasoning is evaluated. A match results in a reward of 1, a similar reasoning results in a reward of 0.5, and a mismatch results in a reward of 0.
* **(3) Multi-view Alignment Reward:** The completion's reasoning is compared to visual evidence.
* Example of a match: `<think>When zoomed in, the eyeball shows structural irregularities, accompanied by high-frequency artifacts...</think>` This is rewarded with R = 1.
* Example of a mismatch: `<think>When zoomed in, the eyeball appears natural, but the high-pass image reveals artifacts...</think>` This is rewarded with R = 0.
### Key Observations
* The diagram highlights the use of MLLMs in determining the authenticity of images.
* The R-GRPO stage uses multiple reward mechanisms to guide the MLLM towards accurate and well-reasoned conclusions.
* The multi-view alignment reward incorporates visual evidence into the reward process.
### Interpretation
The diagram illustrates a sophisticated approach to image authentication using MLLMs and reward-guided policy optimization. The CoE tuning stage likely serves to pre-train the MLLM, while the R-GRPO stage refines the model's reasoning and decision-making process. The use of multiple reward mechanisms, including answer reward, think reward, and multi-view alignment reward, ensures that the MLLM not only provides accurate answers but also generates sound reasoning that aligns with visual evidence. The system aims to mimic human reasoning by considering multiple cues and synthesizing them to arrive at a conclusion. The comparison of zoomed-in views of the eyeball, along with high-pass filtered images, suggests a focus on detecting subtle artifacts that may indicate synthetic image generation.
</details>
Figure 4: Overview of REVEAL. The pipeline mainly consists of two stages: CoE Tuning and R-GRPO.
Chain-of-Evidence Synthesis
As shown in Figure 3, after the specialized expert annotation, the initial eight rounds of multi-perspective diagnostic outputs are diverse and fragmented. To construct a unified and progressive reasoning dataset suitable for Chain-of-Thought (CoT) fine-tuning, we leverage a high-capacity LVLM (Qwen-2.5VL-72B [bai2025qwen2]) to perform structured knowledge consolidation. This process reconstructs the diverse, specialized evidence into a single, cohesive, and auditable reasoning trace, formatted using a standard <think> $·s$ </think> $·$ <answer> $·s$ </answer> structure.
Fundamentally distinct from existing datasets like AIGI-Holmes [zhou2025aigi] and FakeBench [li2025fakebench], which merely provide generic explanations, REVEAL-Bench explicitly formalizes the link between low-level expert evidence and high-level judgments. This two-stage pipeline transforms the detection tasks into a reasoning task, offering coherent CoE annotations that enhance logical consistency, minimizing annotation noise, and support supervision paradigms with advanced reinforcement learning techniques to improve explanation fidelity and generalization.
4 Methodology
4.1 Overview of REVEAL
As illustrated in Figure 4, the overall training pipeline adopts a two-stage progressive training paradigm inspired by advanced policy optimization-based reinforcement learning techniques [guo2025deepseek].
We first perform supervised fine-tuning (SFT) on a consolidated Chain-of-Evidence (CoE) dataset to obtain a base policy that can deduce the required forensic reasoning procedure. While this stage establishes the fundamental reasoning-based forensic structure, the resulting model still exhibits limitations in logical consistency, forensic accuracy, and robustness. To mitigate these limitations, we propose a novel reinforcement learning algorithm: R easoning- e nhanced Forensic E vid e nce A na l ysis (R-GRPO). R-GRPO extends beyond standard Group Relative Policy Optimization (GRPO) by incorporating a task-specific composite reward that dynamically aligns forensic reasoning trajectories and stabilizes policy updates, significantly enhancing semantic consistency and reasoning robustness.
4.2 Progressive Multimodal Training for AI-Generated Image Detection
We introduce REVEAL (Reasoning-enhanced Forensic Evidence AnaLysis), a progressive multimodal training framework comprising two sequential stages designed to cultivate robust, logically consistent, and verifiable forensic reasoning in multimodal models.
Stage 1: Chain-of-Evidence Tuning (CoE Tuning). In the initial stage, we perform cold-start supervised fine-tuning to establish a stable, stepwise reasoning policy and a consistent output paradigm built upon the REVEAL-Bench dataset. Let $x$ denote the visual input, $z=(z_{1},...,z_{T})$ denote the tokenized reasoning sequence (Chain-of-Evidence, CoE), and $y$ denote the final classification label. We adopt an explicit joint reasoning–decision modeling paradigm, where the final prediction $y$ is conditioned on the explicit reasoning trace $z$ . This formulation enforces a think-then-answer mechanism, fundamentally distinct from post-hoc rationalizations (e.g. modeling $p(y\mid x)$ and then $p(z\mid x,y)$ ), thereby achieving causally grounded genuine explanations.
Concretely, we factorize the joint conditional probability as
$$
p(y,z\mid x)\;=\;p(z\mid x)\,p(y\mid x,z), \tag{1}
$$
which structurally encourages the model to first generate verifiable reasoning evidence and subsequently derive the final prediction conditioned directly on that reasoning process.
Maximizing the likelihood under (1) corresponds to minimizing the following negative log-likelihood loss:
$$
\mathcal{L}_{\mathrm{NLL}}(x,y,z;\theta)\;=\;-\log p_{\theta}(z\mid x)\;-\;\log p_{\theta}(y\mid x,z). \tag{2}
$$
For training control and to explicitly balance the emphasis on reasoning quality versus final decision accuracy, we decompose $\mathcal{L}_{\mathrm{NLL}}$ into two components, the reasoning generation loss $\mathcal{L}_{\mathrm{think}}$ and the answer loss $\mathcal{L}_{\mathrm{answer}}$ ,
$$
\mathcal{L}_{\mathrm{think}}\;=\;-\sum_{t=1}^{T}\log p_{\theta}(z_{t}\mid z_{<t},x), \tag{3}
$$
$$
\mathcal{L}_{\mathrm{answer}}\;=\;-\log p_{\theta}(y\mid x,z), \tag{4}
$$
We then employ a weighted composite SFT loss:
$$
\mathcal{L}_{\mathrm{SFT}}=\begin{aligned} &(1-\alpha)\,\mathcal{L}_{\mathrm{think}}+\alpha\,\mathcal{L}_{\mathrm{answer}}+\eta\,\mathrm{KL}\big(\pi_{\mathrm{pre}}\|\pi_{\theta}\big).\end{aligned} \tag{5}
$$
where $\alpha∈(0,1)$ controls the relative importance of the answer loss versus the reasoning trace, the KL regularization term constrains the fine-tuned policy $\pi_{\theta}$ to remain proximal to the pretrained policy $\pi_{\mathrm{pre}}$ , effectively mitigating catastrophic forgetting.
Stage 2: Reasoning-enhanced Group Relative Policy Optimization (R-GRPO).
Group Relative Policy Optimization (GRPO). Group Relative Policy Optimization (GRPO) is a reinforcement learning technique that stabilizes policy updates by comparing a group of candidate trajectories, rather than relying on the noisy reward signals of individual samples. Given an input $x$ , we sample a group of $K$ trajectories $\{\tau_{i}\}_{i=1}^{K}$ from the current policy $\pi_{\theta}$ , where each trajectory $\tau_{i}$ consists of an intermediate reasoning trace $z_{i}$ and a final output $y_{i}$ . A group-based composite reward $R_{\mathrm{group}}(\tau_{i})$ is computed for each trajectory, and the group-relative advantage $A_{i}$ is defined by subtracting the mean group reward $\overline{R}_{\mathrm{group}}$ :
$$
\displaystyle A_{i} \displaystyle=R_{\mathrm{group}}(\tau_{i})-\overline{R}_{\mathrm{group}}, \displaystyle\overline{R}_{\mathrm{group}} \displaystyle=\frac{1}{K}\sum_{j=1}^{K}R_{\mathrm{group}}(\tau_{j}). \tag{6}
$$
The GRPO objective maximizes the expected group-relative log-probability, regularized by a KL penalty for stable policy convergence:
$$
\max_{\theta}\;\mathbb{E}\Big[\sum_{i=1}^{K}A_{i}\log\pi_{\theta}(\tau_{i}\mid x)\Big]\;-\;\lambda_{\mathrm{KL}}\,\mathrm{KL}\big(\pi_{\mathrm{old}}\|\pi_{\theta}\big). \tag{7}
$$
Reasoning-enhanced GRPO (R-GRPO). To employ GRPO for forensic analysis tasks, we propose R-GRPO, which augments the objective with a task-aware composite reward specifically designed to capture forensic fidelity and reasoning robustness. Let $y$ denote the generated answer, $y^{\ast}$ the reference answer, $z=(z_{1},...,z_{T})$ the reasoning tokens, and $\{v_{m}(x)\}_{m=1}^{M}$ a set of multi-visual visual evidence (e.g., spectral representations, high-pass filtered images, and localized artifact patches).
Rationale for Agent-based Reward Modeling. In preliminary experiments, we observed that simple metric-based rewards (e.g. using cosine similarity of sentence embeddings for $r_{\mathrm{sem}}$ ) fail to adequately reflect the semantic and contextual logic required for high-quality forensic explanations. Therefore, we introduce a dedicated large language model as an intelligent agent (Agent) to evaluate responses. This Agent-based assessment considers contextual logic, explanation coherence, and factual consistency against the provided structured evidence, thereby generating a more human-aligned and interpretable reward signal than purely metric-based approaches (see Appendix A for details).
R-GRPO defines three complementary, evidence-driven reward components:
(1) Answer Reward $r_{\mathrm{sem}}$ . This binary reward ensures the accuracy of the detection:
$$
r_{\mathrm{sem}}(y,y^{\ast})=\begin{cases}1,&\text{if }y=y^{\ast},\\
0,&\text{otherwise.}\end{cases} \tag{8}
$$
(2) Think Reward $r_{\mathrm{think}}$ . This reward quantifies the quality and structural integrity of the reasoning trace $z$ .
Let $z=(z_{1},...,z_{T})$ be the generated reasoning trace and $z^{\ast}=(z^{\ast}_{1},...,z^{\ast}_{T^{\ast}})$ the ground-truth reasoning trace (when available). Define a perturbed trace $\tilde{z}=\operatorname{shuffle}(z)$ . Then
$$
r_{\mathrm{think}}(z,z^{\ast},\tilde{z})\;=\;\mathcal{A}_{\mathrm{sem}}(z,z^{\ast})+\mathcal{A}_{\mathrm{logic}}(z,\tilde{z}), \tag{9}
$$
where $\mathcal{A}_{\mathrm{sem}}$ measures alignment between the generated and reference reasoning, and $\mathcal{A}_{\mathrm{logic}}(z,\tilde{z})$ evaluates the logical coherence of the trace. Crucially, $\mathcal{A}_{\mathrm{logic}}$ evaluates the logical coherence by penalizing the model if minor structural perturbations $\tilde{z}$ severely alter the inferred conclusion. This mechanism forces the model to maintain sequential consistency and ensure the reasoning steps are robustly connected.
(3) Multi-view Alignment Reward $r_{\mathrm{view}}$ . This reward encourages the generated reasoning trace $z$ to be robustly grounded in evidence that persists across different forensic views of the image.
$$
r_{\mathrm{view}}(z,x)\;=\;\mathcal{A}_{\mathrm{view}}\Big(z,\{v_{m}(x)\}_{m=1}^{M}\Big), \tag{10}
$$
where $\mathcal{A}_{\mathrm{view}}$ measures fidelity of the reasoning to the multi-view visual evidence $\{x_{m}\}$ . By requiring alignment with evidence visible under different transformations (e.g., spectral, high-pass), this reward promotes cross-artifact generalization and enables the self-supervised discovery of novel, transformation-invariant artifacts.
The composite trajectory reward $R(\tau)$ combines these terms:
$$
\displaystyle R(\tau)= \displaystyle\lambda_{s}r_{\mathrm{sem}}(y,y^{\ast})+\lambda_{t}r_{\mathrm{think}}(z,z^{\ast},\tilde{z}) \displaystyle+\lambda_{v}r_{\mathrm{view}}(z,x), \tag{11}
$$
where $\lambda_{s},\lambda_{t},\lambda_{v}≥ 0$ are tunable parameters balancing the rewards. For improved stability, rewards are standardized within each sampled group before calculating the advantage $\widehat{A}_{i}$ :
$$
\widehat{R}(\tau_{i})=\frac{R(\tau_{i})-\mu_{\mathrm{group}}}{\sigma_{\mathrm{group}}}, \tag{12}
$$
$$
\mu_{\mathrm{group}}=\frac{1}{K}\sum_{j}R(\tau_{j}), \tag{13}
$$
$$
\sigma_{\mathrm{group}}=\mathrm{std}(\{R(\tau_{j})\}) \tag{14}
$$
and the normalized group-relative advantage is
$$
\widehat{A}_{i}=\widehat{R}(\tau_{i})-\frac{1}{K}\sum_{j}\widehat{R}(\tau_{j}). \tag{15}
$$
Unified GRPO with the R-GRPO objective. Combining the original GRPO formulation (7) with the R-GRPO composite reward (11), the unified optimization objective becomes
$$
\max_{\theta}\;\mathbb{E}\Big[\sum_{i=1}^{K}\widehat{A}_{i}\log\pi_{\theta}(\tau_{i}\mid x)\Big]\;-\;\lambda_{\mathrm{KL}}\,\mathrm{KL}\big(\pi_{\mathrm{old}}\|\pi_{\theta}\big), \tag{16}
$$
where $\widehat{A}_{i}$ encodes both the group-relative comparison and the reasoning-enhanced composite reward.
This evidence-enhanced reward signals can effectively guide the model to optimize its reasoning trajectories, enforcing both stability and logical coherence in verifiable forensic evidence analysis.
5 Experiments
5.1 Experimental Settings
TABLE I: Comparison of REVEAL-bench with previous datasets. REVEAL-bench is the first reasoning dataset for synthetic image detection.
Dataset #Image Explanation Multiview Fusion Reasoning Process CNNDetection [wang2020cnn] 720k ✗ ✗ ✗ GenImage [zhu2023genimage] 1M ✗ ✗ ✗ FakeBench [li2025fakebench] 6K ✓ ✗ ✗ Holmes-Set [zhou2025aigi] 69K ✓ ✓ ✗ REVEAL-bench 60K ✓ ✓ ✓
To comprehensively evaluate the performance of REVEAL, we conduct experiments on two datasets: REVEAL-Bench and GenImage [zhu2023genimage] (see Table I). REVEAL-Bench, the first chain-of-evidence-based explainable dataset for synthetic image detection, serves as the in-domain dataset for training and evaluation. GenImage, a large-scale synthetic image dataset containing images generated by multiple generation methods, is used as an out-of-domain dataset to assess generalization. We train REVEAL on REVEAL-Bench and systematically evaluate its performance on both datasets (see Appendix B for detailed training settings). Building on this evaluation setup, we further investigate several core aspects of REVEAL’s capabilities. In particular, we study the impact of different MLLMs used as vision–language backbones, conduct ablation experiments to quantify the contribution of R-GRPO, and assess the model’s robustness under diverse perturbation settings. Appendix C reports the few-shot training results, and Appendix D provides a systematic comparison with existing large-scale model-based detectors.
Baselines We compare REVEAL with state-of-the-art AI-generated image detection methods, including CNNSpot [wang2020cnn], UnivFD [ojha2023towards], NPR [tan2024rethinking], HyperDet [cao2024hyperdet], AIDE [yan2024sanity] and VIB-Net [zhang2025towards]. To ensure a fair comparison, we retrain these methods using the official code under the same experimental settings and datasets.
Evaluation metrics Following existing research, we adopt Accuracy (ACC) as our evaluation metric. Accuracy is defined as the proportion of correctly predicted samples among the total number of samples, reflecting the overall correctness of a classification model. Since our detection results are provided by the MLLM in textual form (Real/Fake), we convert these texts into binary labels to compute accuracy, while baseline methods use the default thresholds provided by their official code. Moreover, because the output of the MLLM is interpretable text rather than logit values, we do not consider metrics that require logit values for computation, such as Average Precision (AP), in our evaluation.
5.2 Generalization across datasets
TABLE II: REVEAL demonstrates superior generalization across both in-domain and out-of-domain evaluations. REVEAL outperforms the best competing method by 3.87 %.
| CNNSpot [wang2020cnn] UnivFD [ojha2023towards] NPR [tan2024rethinking] | 87.80 86.95 95.40 | 62.45 75.00 84.80 | 74.25 84.35 88.85 | 73.85 80.95 88.05 | 63.55 85.50 85.10 | 73.60 71.75 94.30 | 73.70 82.00 87.05 | 71.35 80.70 84.45 | 39.45 88.45 88.95 | 68.89 81.74 88.55 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| HyperDet [cao2024hyperdet] | 93.25 | 68.40 | 91.85 | 92.30 | 100.0 | 67.05 | 89.20 | 80.45 | 57.65 | 82.24 |
| AIDE [yan2024sanity] | 95.25 | 79.90 | 95.90 | 94.95 | 87.75 | 90.35 | 94.85 | 90.10 | 91.10 | 91.13 |
| VIB-Net [zhang2025towards] | 67.05 | 53.25 | 60.25 | 57.85 | 65.00 | 68.55 | 60.85 | 52.55 | 38.00 | 58.15 |
| REVEAL | 95.31 | 93.75 | 97.81 | 97.19 | 95.00 | 86.88 | 96.25 | 95.94 | 96.88 | 95.00 |
Table II reports the performance of REVEAL on the in-domain dataset REVEAL-bench and the out-of-domain benchmark GenImage. The results indicate that REVEAL, leveraging a Chain-of-Evidence (CoE) reasoning-and-forensics mechanism, achieves superior cross-domain generalization compared to baseline lightweight binary classifiers: it maintains higher accuracy and more stable performance on GenImage. In the in-domain setting, smaller classifiers, such as those using methods like NPR [tan2024rethinking] and AIDE [yan2024sanity], are more prone to overfitting, demonstrating stronger fitting ability to domain-specific statistical regularities and subtle signals. As a result, REVEAL’s performance in-domain is comparable to that of these compact models. However, REVEAL excels in terms of cross-domain generalization. These findings suggest that while smaller models remain attractive for tasks prioritizing computational efficiency and in-domain accuracy, REVEAL better preserves and propagates key reasoning cues across domains. Therefore, there is a clear trade-off between generalization and domain-specific fit that should inform deployment choices. Notably, in the context of synthetic-image detection, reasoning-based forensic approaches, like REVEAL, exhibit particularly robust generalization.
5.3 Generalization across Base MLLMs
TABLE III: Performance across different MLLMs, showing larger models exhibit consistently stronger detection capability.
Training Scheme Phi-3.5 Qwen2.5- VL-3b Qwen2.5- VL-7b llava- v1.5-7b llava- v1.5-13b CoE Tuning 83.75 87.18 85.73 91.56 93.06 CoE Tuning+G-GRPO 87.19 89.06 92.19 92.81 95.31
The proposed algorithm in this study demonstrates strong generalizability and can be flexibly applied to a variety of multimodal large model architectures. To validate the effectiveness of our method, we conduct experiments using Qwen2.5-VL [bai2025qwen2], LLaVA-1.5-VL [liu2023visual], and Phi-3.5 as representative training frameworks. As shown in Table III, the results indicate that our approach achieves excellent detection performance and robust generalization across different multimodal large models.
Furthermore, we observe that as the model size increases, the detection capability improves significantly. This trend suggests the existence of a scaling law for synthetic image detection within the context of large models, similar to other tasks in the large model domain. As multimodal models continue to grow, their ability to handle complex tasks such as synthetic image detection becomes increasingly effective, demonstrating a direct correlation between model scale and performance.
5.4 Ablation Studies
TABLE IV: Ablation study of the impact of CoE Tuning, GRPO, and R-GRPO on model accuracy on REVEAL-Bench.
| ✗ | ✗ | ✗ | 61.21 |
| --- | --- | --- | --- |
| ✓ | ✗ | ✗ | 85.73 |
| ✓ | ✓ | ✗ | 91.56 |
| ✓ | ✗ | ✓ | 95.31 |
We conducted ablation experiments to investigate the role of reasoning datasets in synthetic image detection. As shown in Table IV, we first evaluated the performance of models trained without reasoning data (i.e., non-Reasoning SFT) and compared them with models fine-tuned using reasoning data (i.e., CoE Tuning). Additionally, we tested the effects of applying simple GRPO and our proposed R-GRPO method on performance improvement. The experimental results demonstrate that reasoning datasets significantly enhance the performance of MLLMs in synthetic image detection, with models lacking reasoning data performing close to random levels. Moreover, applying G-GRPO further improved the performance, highlighting the critical role of R-GRPO in this task.
5.5 Robustness Evaluation of REVEAL
<details>
<summary>x5.png Details</summary>
### Visual Description
## Line Charts: REVEAL-bench Performance
### Overview
The image contains two line charts comparing the performance of "Ours(REVEAL)" and "NPR(2024CVPR)" on the REVEAL-bench dataset. The left chart shows accuracy (Acc) vs. quality, while the right chart shows accuracy vs. sigma. A horizontal dashed line is present at Acc = 50 in both charts.
### Components/Axes
**Legend:** Located at the top of the image.
* Red line with circle markers: Ours(REVEAL)
* Blue line with triangle markers: NPR(2024CVPR)
**Left Chart:**
* Title: REVEAL-bench
* Y-axis: Acc (Accuracy), ranging from 50 to 100 in increments of 10.
* X-axis: quality, ranging from 60 to 100 in increments of 10.
**Right Chart:**
* Title: REVEAL-bench
* Y-axis: Acc (Accuracy), ranging from 50 to 100 in increments of 10.
* X-axis: sigma, ranging from 0 to 4 in increments of 1.
### Detailed Analysis
**Left Chart (Accuracy vs. Quality):**
* **Ours(REVEAL) (Red):** The accuracy decreases as quality decreases.
* Quality 100: Acc ~ 95
* Quality 90: Acc ~ 77
* Quality 80: Acc ~ 65
* Quality 70: Acc ~ 60
* Quality 60: Acc ~ 58
* **NPR(2024CVPR) (Blue):** The accuracy is relatively flat and low as quality decreases.
* Quality 100: Acc ~ 57
* Quality 90: Acc ~ 52
* Quality 80: Acc ~ 51
* Quality 70: Acc ~ 50
* Quality 60: Acc ~ 50
**Right Chart (Accuracy vs. Sigma):**
* **Ours(REVEAL) (Red):** The accuracy decreases as sigma increases, then flattens out.
* Sigma 0: Acc ~ 95
* Sigma 1: Acc ~ 81
* Sigma 2: Acc ~ 66
* Sigma 3: Acc ~ 60
* Sigma 4: Acc ~ 58
* **NPR(2024CVPR) (Blue):** The accuracy decreases as sigma increases, then flattens out.
* Sigma 0: Acc ~ 95
* Sigma 1: Acc ~ 81
* Sigma 2: Acc ~ 58
* Sigma 3: Acc ~ 55
* Sigma 4: Acc ~ 54
### Key Observations
* In both charts, "Ours(REVEAL)" consistently outperforms "NPR(2024CVPR)".
* The accuracy of "Ours(REVEAL)" is more sensitive to changes in quality and sigma than "NPR(2024CVPR)".
* "NPR(2024CVPR)" performance is close to the 50 accuracy baseline for lower quality values.
### Interpretation
The charts demonstrate that the "Ours(REVEAL)" method is more robust to variations in both quality and sigma compared to "NPR(2024CVPR)" on the REVEAL-bench dataset. The significant drop in accuracy for "Ours(REVEAL)" as quality decreases or sigma increases suggests that these parameters have a substantial impact on its performance, but it still maintains a higher accuracy than "NPR(2024CVPR)". The relatively flat performance of "NPR(2024CVPR)" indicates that it is less affected by these parameters, but its overall accuracy is lower. The horizontal line at 50 accuracy likely represents a baseline or threshold, suggesting that "NPR(2024CVPR)" approaches this baseline under certain conditions.
</details>
Figure 5: The accuracy comparison between the two methods under various perturbation conditions.
To evaluate the robustness of REVEAL against common post-processing distortions, we conducted a systematic robustness study on the REVEAL-bench dataset. The experiments apply two typical post-processing operations to the original test images: Gaussian blur ( $\sigma=1,2,3,4$ ) and JPEG compression (quality = 90, 80, 70, 60). For each distortion level, we compare REVEAL with the state-of-the-art baseline methods (results are shown in Figure 5). The results indicate that REVEAL demonstrates stronger robustness and improved cross-domain generalization across the considered post-processing settings.
6 Conclusion
We presented REVEAL, a reasoning-centered approach for explainable AI-generated image detection. First, we introduced REVEAL-Bench, the first dataset organized around expert-grounded, verifiable forensic evidence and an explicit chain-of-evidence following an evidence-then-reasoning paradigm. Second, we proposed the REVEAL Framework, a progressive two-stage training scheme whose core component R-GRPO explicitly teaches multimodal LLMs to perform logical synthesis over forensic evidence, jointly improving accuracy, reasoning consistency, and generalization. Empirically, REVEAL attains superior detection accuracy, stronger out-of-domain generalization, and higher explanation fidelity, establishing a new state of the art for reasoning-based image forensics.