ECCV 2026 Multimodal VLM AI-generated image detection chain-of-evidence reasoning GRPO reinforcement learning explainable forensics cross-generator generalization

REVEAL: Reasoning-Enhanced Forensic Evidence Analysis for Explainable AI-Generated Image Detection¶

Conference: ECCV 2026
Paper: ECCV
Code: https://github.com/TrustMedia-zju/REVEAL
Area: Multimodal VLM
Keywords: AI-generated image detection, chain-of-evidence reasoning, GRPO reinforcement learning, explainable forensics, cross-generator generalization

TL;DR¶

REVEAL first uses eight lightweight expert models to extract structured forensic evidence, has a large model consolidate it into chain-of-evidence (CoE) annotations, and distills a 60K REVEAL-Bench; it then trains an MLLM with a two-stage recipe (CoE tuning followed by R-GRPO with an evidence-grounded composite reward), reaching 95.31% in-domain and 94.96% cross-domain on GenImage while producing explanations that can be checked against low-level evidence.

Background & Motivation¶

AI-generated images are now realistic enough to fool human perception, and diffusion models such as FLUX and SDv3.5 keep raising the bar, which makes reliable detection an urgent priority. Existing detectors fall into two groups, and both offer limited forensic value. The first is purely discriminative classifiers: they evolved from hand-crafted low-level statistics (up-sampling grid artifacts, high-frequency decay, spectral anomalies) to CNN/ViT detectors that learn discriminative features directly (CNNSpot, UnivFD, NPR, HyperDet, AIDE, VIB-Net). They emit a label and nothing else. The second group is somewhat more interpretable — rule-based models and decision trees expose part of the decision — but it generalizes poorly (in this paper's Table 6, majority voting over the eight experts reaches only 78.35% and a decision tree 74.75%), and its judgments remain tightly coupled to the feature biases of the underlying detector.

Multimodal large language models look like the natural route to explainability (GPT-4-based detection, AIGI-Holmes, FakeBench, RAIDX), but they suffer from two structural problems. First, their explanations are mostly post hoc: the model computes \(p(y\mid x)\), commits to a verdict, and then appends a rationale \(p(z\mid x,y)\), with no mechanism ensuring that the intermediate steps are actually supported by verifiable evidence. Second, the MLLM is used as a general-purpose visual classifier that latches onto high-level anomalies ("unnatural lighting", "blurred boundaries") rather than as a forensic system that systematically collects, analyzes, and synthesizes specialized evidence. The deeper cause is the absence of suitable data and training objectives: existing benchmarks provide either image-level labels or a single short textual justification (e.g. FakeBench), and neither carries the structured evidence forensic analysis requires, while retrieval-augmented prompting such as RAIDX produces fluent explanations that never bind to image-specific evidence chains.

What is genuinely missing, then, is two things — a reasoning-oriented forensic dataset whose annotations contain structured evidence and step-by-step support for the final judgment, and a training objective that enforces evidence grounding. The answer given here is to extract low-level evidence offline with eight public lightweight experts (artifact masks plus diagnostic labels), let Qwen-2.5VL-72B consolidate them into a single CoE trace to obtain a 60K REVEAL-Bench, and then write that evidence chain into the MLLM's policy with a two-stage recipe: CoE tuning to learn the canonical structure, then R-GRPO with a composite reward (answer correctness, reasoning stability, multi-view consistency). Core idea: reverse the modeling order from "decide first, rationalize afterwards" to "collect verifiable evidence, reason along the evidence chain, then decide", and train that chain into the policy with three complementary rewards, so that explainability becomes the source of generalization rather than a post-hoc decoration.

Method¶

Overall Architecture¶

REVEAL consists of three pieces: the offline REVEAL-Bench, the Stage-1 CoE tuning, and the Stage-2 R-GRPO. The input is an image to be inspected (plus a "real or generated?" question); the output is a binary verdict together with a <think>…</think><answer>…</answer> evidence chain. The critical premise is that the eight expert models are used only during offline dataset construction to produce artifact annotations and diagnostic labels — they participate in neither training nor inference, so deployment costs no more than an ordinary MLLM detector. On the data side, six public detection benchmarks are pooled into roughly 5,120K synthetic and 850K authentic images, stratified-sampled by aesthetic score and resolution, annotated by the eight experts, and finally consolidated by Qwen-2.5VL-72B into a 60K CoE corpus (30K real, 30K synthetic). On the training side, a backbone such as llava-v1.5-13B first undergoes CoE tuning to establish a stable think-then-answer output convention, and is then optimized by R-GRPO with group-relative advantages over three reward terms.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Expert-grounded evidence<br/>and CoE construction"] --> B["CoE Tuning<br/>joint think-then-answer modeling"]
    B --> C["R-GRPO<br/>evidence-grounded composite reward"]
    C --> D["REVEAL detector<br/>verdict + verifiable explanation"]

Key Designs¶

1. Expert-grounded evidence and CoE construction: making annotation cite evidence instead of inventing rationales

The pain point on the data side is that image-level labels or a free-form sentence cannot support forensic analysis, while the diagnostics of several experts, taken separately, are fragmented and view-specific. The first step pools six public benchmarks (CNNDetection, UnivFD, AIGCDetectBenchmark, GenImage, Fake2M, Chameleon) into roughly 5,120K synthetic and 850K authentic images, then stratified-samples by aesthetic score (50% high, 30% medium, 20% low) and resolution (50% at ≥512×512, 30% between 384×384 and 512×512, 20% below 384×384), covering 13 semantic categories so that the corpus does not over-concentrate on particular generators or quality levels. The second step has eight lightweight experts screen one artifact family each: local artifacts, spectral clues, pixel noise (NPR), spatial consistency, geometry flaws, shadow logic, texture fusion, and high-pass fusion. All of them are drawn from public codebases and adopted without any additional training, and what they emit is machine-readable structured evidence — artifact masks plus diagnostic labels — rather than a vague "this looks fake". The last step uses Qwen-2.5VL-72B to consolidate the eight multi-view diagnostics into one coherent, auditable reasoning trace in the unified <think>…</think><answer>…</answer> format; from a pool of roughly 6M annotated samples, stringent filtering distills 60K (half real, half synthetic), a selection rate of about 1%.

What truly separates this from AIGI-Holmes is how the experts are used: there, experts perform global image filtering and their conclusions never enter the explanation; here, experts emit structured low-level evidence, and the causal chain "low-level artifact → intermediate inference → final conclusion" is written explicitly into the annotation. Explanations trained this way can be checked against evidence, and the model no longer learns pseudo-correlations of the form "images of this kind usually look like that".

2. CoE tuning: writing "reason first, judge later" into the probability factorization

An MLLM trained with ordinary SFT still treats the explanation as an appendage of the verdict, with no causal constraint between the reasoning chain and the conclusion. Stage 1 therefore does more than feed CoT data: it explicitly factorizes the joint probability as \(p(y,z\mid x)=p(z\mid x)\,p(y\mid x,z)\) — generate the evidence chain \(z\) first, then produce the verdict \(y\) conditioned on \(z\). This looks like a mere reordering of a product, but it differs fundamentally from the prior "compute \(p(y\mid x)\) then append \(p(z\mid x,y)\)" formulation. The latter lets the model follow its gut, commit, and then reverse-engineer a plausible-sounding rationale (exactly post-hoc rationalization); the former requires the verdict to appear only after the reasoning chain has been generated, so every conclusion must travel along evidence that has already been written down.

To keep training controllable, the negative log-likelihood is decomposed into a reasoning-generation term and an answer term that are combined with weights, plus a KL regularizer that pulls the fine-tuned policy \(\pi_\theta\) back toward the pretrained policy \(\pi_{\text{pre}}\) to mitigate catastrophic forgetting. The weight \(\alpha\) is set to 0.1, so reasoning serves as an auxiliary signal while answer correctness remains the primary target. The value of this stage shows up plainly in the ablations: SFT with no reasoning supervision reaches only 61.21%, an answer-first reasoning format reaches 82.39%, and a single CoE tuning step brings it to 85.73% — evidence that what matters is not merely the presence of reasoning text but the structural constraint that the chain precedes the answer.

3. R-GRPO: turning reasoning quality into an optimizable signal via an evidence-grounded composite reward

CoE tuning settles the question of format, but not whether the reasoning chain is genuinely stable and genuinely about this image. On top of the group-relative advantage of standard GRPO, R-GRPO substitutes a forensics-specific composite reward with three complementary components. The answer reward \(r_{\text{ans}}\) is the simplest binary term, rewarding agreement with the ground-truth verdict. The think reward \(r_{\text{think}}\) has two parts: one measures semantic alignment between the generated chain and the reference chain, and the other measures the chain's logical coherence — the latter with a neat trick. The chain is lightly perturbed (shuffled) into \(\tilde{z}\); if such a perturbation flips the inferred conclusion, the original reasoning was unstable and is penalized. The multi-view alignment reward \(r_{\text{view}}\) requires the content inside <think> to match what the same image shows under several forensic views (spectral representation, high-pass filtered image, edge responses, localized artifact regions): every view must be analyzed correctly and the views must not contradict one another. This constrains explanation faithfulness and also encourages generalization across artifact types.

The three terms are combined by a weighted sum \(R(\tau)=\lambda_a r_{\text{ans}}+\lambda_t r_{\text{think}}+\lambda_v r_{\text{view}}\), with \((\lambda_a,\lambda_t,\lambda_v)=(0.8,0.1,0.1)\) so that the reasoning and view rewards only support the verdict rather than dominate it. One easily overlooked but crucial implementation choice is that none of the three — semantic alignment, logical coherence, view alignment — is computed by a metric such as sentence-embedding cosine similarity; instead Qwen-3-VL-8B serves as a judge agent requiring no additional training. In preliminary experiments the authors found that purely metric-based rewards cannot tell whether the contextual logic holds at all; only a vision-language model reading the response against the actual evidence yields a reward signal better aligned with human judgment. Precisely because the reward is read rather than computed, it can look at the reasoning chain and the image evidence at the same time, pressing both "it makes sense" and "it matches the image" into a single optimization target.

A Worked Example¶

Take the synthetic bird image in Figure 2 of the paper. The eight experts return view-specific conclusions: the local-artifact expert, zooming in, points out that the pupil region lacks a natural reflection; the spectral expert finds periodic peaks along the frequency axis typical of synthesized images; the high-pass fusion expert notices that the area around the bird is suspiciously smooth with no sign of forgery at all. These conclusions do not agree — one of them even considers the image real. The job of CoE synthesis is to merge them into one prioritized seven-step trace: an initial screening of texture and lighting anomalies (1), a closer inspection of uniform surfaces and missing imperfections (2), a comparison of object–background spatial alignment and projective logic (3), a shadow-consistency check (4), a high-frequency texture-irregularity analysis (5), an evaluation of abnormal spectral energy distribution (6), and finally a synthesis of all cues into a synthetic verdict (7) with <answer>1</answer>. The verdict is not produced by any single expert's vote but merged from six mutually independent technical cues — which is exactly the difference between 78.35% for majority voting and 95.31% for REVEAL.

Loss & Training¶

Stage 1 is cold-start supervised fine-tuning. Taking the negative log-likelihood of the factorization yields the NLL loss, which the paper decomposes into a reasoning-generation term and an answer term, combined with weights and a KL regularizer:

\[\mathcal{L}_{\text{SFT}} = \alpha\,\mathcal{L}_{\text{think}} + (1-\alpha)\,\mathcal{L}_{\text{answer}} + \beta\,\mathrm{KL}(\pi_{\text{pre}} \| \pi_\theta)\]

⚠️ Equation (5) of the original paper is corrupted in the cache, so the symbols are reconstructed to match the prose: \(\alpha=0.1\) keeps reasoning auxiliary and prioritizes answer correctness, and the KL term suppresses catastrophic forgetting. Refer to the original paper.

Stage 2 is R-GRPO. It follows GRPO in replacing the critic with a group-based baseline: for a given input, \(G\) trajectories \(\{\tau_i\}\) are sampled and the composite reward is standardized within the group to serve directly as the advantage,

\[\hat{A}_i = \frac{R(\tau_i) - \mu_{\text{group}}}{\sigma_{\text{group}}},\]

which is then plugged into a PPO-style surrogate objective, with clipping on the probability ratio and a KL penalty anchoring the policy to a reference policy \(\pi_{\text{ref}}\) (typically the SFT model):

\[\mathcal{J}_{\text{R-GRPO}}(\theta)=\mathbb{E}_{x}\frac{1}{G}\sum_{i=1}^{G}\min\!\Big(\frac{\pi_\theta(\tau_i\mid x)}{\pi_{\text{old}}(\tau_i\mid x)}\hat{A}_i,\ \mathrm{clip}\big(\tfrac{\pi_\theta}{\pi_{\text{old}}},1-\epsilon,1+\epsilon\big)\hat{A}_i\Big) - \beta\,D_{\mathrm{KL}}(\pi_\theta \| \pi_{\text{ref}})\]

⚠️ This objective (Equation (13) of the original) is likewise corrupted in the cache; the structure is reconstructed in the standard GRPO form. Refer to the original paper. Key hyper-parameters: SFT weight \(\alpha=0.1\), reward weights \((\lambda_a,\lambda_t,\lambda_v)=(0.8,0.1,0.1)\), Qwen-3-VL-8B as the reward judge, and Qwen-2.5VL-72B for CoE synthesis. Note also that the paper expands the R-GRPO acronym inconsistently ("Reasoning-enhanced Group Relative Preference Optimization" vs. "Reasoning-enhanced Forensic Evidence Analysis"); refer to the original paper.

Key Experimental Results¶

Main Results¶

Evaluation covers three sets: the in-domain REVEAL-Bench, the out-of-domain GenImage, and the newly built REVEAL-Bench++ (10K images, 2K each from FLUX, FLUX2, Z-Image, Qwen-Image, and SDv3.5, half real and half generated, none of the five generators seen in training). All baselines are retrained with their official code under identical splits and protocols.

Method	REVEAL-Bench	SDv3.5	FLUX	FLUX2	Qwen-Image	Z-Image	Mean
CNNSpot	87.80	71.70	70.30	61.70	82.35	59.10	72.16
UnivFD	86.95	84.45	84.55	83.65	85.85	66.75	82.03
NPR	95.40	53.00	51.20	51.60	51.40	53.60	59.37
HyperDet	93.25	88.80	79.70	65.20	88.15	70.80	80.98
AIGI-Holmes	93.10	82.14	79.35	76.41	75.43	69.47	79.32
Veritas	72.75	87.35	89.90	90.76	90.00	87.80	86.43
Ivy-xDetector	77.65	90.75	91.60	91.04	91.65	86.55	88.21
REVEAL	95.31	94.38	93.44	91.25	95.00	84.69	92.35

(Accuracy %, the Mean column averaging the six preceding columns as in the original. REVEAL improves over the strongest baseline, Ivy-xDetector, by 4.14 points.)

Method	Midjourney	SD v1.4	SD v1.5	ADM	GLIDE	Wukong	VQDM	BigGAN	Mean
CNNSpot	62.45	74.25	73.85	63.55	73.60	73.70	71.35	39.45	66.53
UnivFD	75.00	84.35	80.95	85.50	71.75	82.00	80.70	88.45	81.09
NPR	84.80	88.85	88.05	85.10	94.30	87.05	84.45	88.95	87.69
HyperDet	68.40	91.85	92.30	100.0	67.05	89.20	80.45	57.65	80.86
AIDE	79.90	95.90	94.95	87.75	90.35	94.85	90.10	91.10	90.61
VIB-Net	53.25	60.25	57.85	65.00	68.55	60.85	52.55	38.00	57.04
AIGI-Holmes	86.10	93.17	91.22	84.32	72.53	92.10	89.77	91.00	87.53
REVEAL	93.75	97.81	97.19	95.00	86.88	96.25	95.94	96.88	94.96

(Cross-domain generalization on GenImage, accuracy %; REVEAL beats the previous best by at least 4.35 points.)

Ablation Study¶

Training-strategy ablations (Table 5) and backbone swaps (Table 4), both on REVEAL-Bench:

Config	Accuracy (%)	Note
Non-reasoning SFT	61.21	no reasoning data or supervision at all
Answer-first format	82.39	reasoning text present, but the answer precedes it
CoE Tuning	85.73	CoE data and the canonical structure
CoE Tuning + GRPO	91.56	vanilla GRPO
CoE Tuning + R-GRPO	95.31	the full two-stage recipe

⚠️ The original table does not state which backbone was used; the numbers 85.73 and 91.56 coincide with the CoE results for Qwen2.5-VL-7B and LLaVA-1.5-7B in Table 4. Refer to the original paper for the backbone attribution.

Backbone	CoE Tuning	+ R-GRPO
Phi-3.5	83.75	87.19
Qwen2.5-VL-3B	87.18	89.06
Qwen2.5-VL-7B	85.73	92.19
LLaVA-1.5-7B	91.56	92.81
LLaVA-1.5-13B	93.06	95.31

(Accuracy %; R-GRPO improves every backbone, and larger backbones end up better.)

Method	Accuracy (%)
Best lightweight expert	65.48
Majority voting over eight experts	78.35
Decision tree	74.75
REVEAL	95.31

(Table 6: against rule-based aggregation of the same eight expert outputs, showing that synthesizing expert predictions is far weaker than having a large model reason along the evidence chain.)

Key Findings¶

Reasoning supervision delivers the largest single jump, but the format is not the point. SFT without reasoning data reaches only 61.21%; adding reasoning text in an answer-first format gets to 82.39%; restructuring it into the CoE think-then-answer form reaches 85.73%. With the same amount of reasoning text, reversing the order costs more than three points, so the structural constraint that the chain precedes the verdict matters as much as the data itself.
The gain from R-GRPO over vanilla GRPO is nearly as large as going from no reasoning to reasoning at all. GRPO 91.56% → R-GRPO 95.31%, a 3.75-point improvement, and it appears consistently across Phi-3.5, Qwen2.5-VL-3B/7B and LLaVA-1.5-7B/13B — evidence that the evidence-grounded composite reward is model-agnostic rather than a tuning windfall for one backbone.
The narrow in-domain margin and the wide cross-domain margin are the most notable empirical pattern. NPR actually reaches 95.40% on REVEAL-Bench, marginally above REVEAL's 95.31%, yet averages only 59.37% on REVEAL-Bench++ with 51–53% on FLUX, Qwen-Image, and Z-Image; REVEAL stays between 84.69% and 95.00% on all five unseen generators. Compact models are good at fitting in-domain statistical regularities and collapse under distribution shift, whereas the payoff of evidence-chain reasoning shows up mainly on generators never seen before.
REVEAL also stays ahead under perturbation. With Gaussian blur (\(\sigma=4\)) and JPEG compression (quality = 90/80/70/60) applied to test images, REVEAL remains more stable than the baselines at every severity level.
Explanation quality is assessed with several complementary probes rather than impressions. A human preference study (Appendix E) reports that REVEAL's explanations are preferred over prior methods 26% more often; multi-view faithfulness (Appendix G) and a comparison against both closed- and open-source interpretability approaches (Appendix H) corroborate this. On the data side, Figure 4 compares expert-guided against purely LLM-generated annotations along three axes — annotation correctness, description accuracy on GAN images known to carry frequency artifacts, and a human review of 100 samples containing specialized forensic terms — and expert guidance wins on all three. It is worth noting that the main tables report accuracy only: because REVEAL emits textual verdicts rather than calibrated logits, the paper deliberately omits score-based metrics such as average precision.

Highlights & Insights¶

"Experts are used only to build data, never in training or inference" is a pragmatic engineering trade-off. It reduces deployment to an ordinary MLLM detector with no eight-model inference graph to maintain, while preserving the supervision those experts provide — effectively distilling the perception of eight small models into a large model's reasoning chain.
Reducing the gap between post-hoc rationalization and causally grounded explanation to an ordering of a probability factorization (\(p(z\mid x)p(y\mid x,z)\) vs. \(p(y\mid x)p(z\mid x,y)\)) is the paper's "aha" moment. It turns the philosophical-sounding question "is this explanation faithful?" into a modeling-order constraint that can be enforced at the loss level.
Using a shuffled reasoning chain to test logical stability — does the conclusion flip under a light perturbation? — is a cheap and effective reward that needs almost no extra annotation. The idea transfers to any task where reasoning-chain quality is hard to measure directly: apply an equivalence-preserving transformation to intermediate steps in math reasoning, or an irrelevant reordering of statements in code generation, and check whether the conclusion survives.
Using an untrained large model as a reward judge instead of a metric like embedding cosine similarity fits subjective dimensions such as explanation quality particularly well; the cost is one extra large-model forward pass per evaluation, which has to be weighed against throughput.
The multi-view alignment reward operationalizes "does the explanation match the image?" as "does it hold under spectral, high-pass, edge, and localized-artifact views" — enumerable views are an easy recipe to port to other forensic tasks such as video forgery detection or document tampering localization.

Limitations & Future Work¶

The main experiments report accuracy only; because the model emits text rather than logits, score-based metrics such as average precision are absent, and the mapping from textual Real/Fake to a binary label makes the comparison with logit-based detectors not entirely symmetric.
The ablation table (Table 5) does not state the backbone, and its numbers coincide with those of different backbones in Table 4, so a reader cannot be fully certain which model size each row corresponds to; this weakens the quantitative claim about how much reasoning supervision and R-GRPO each contribute.
Although the annotation pool is described as roughly 6M samples, the final corpus is 60K (about a 1% selection rate) distilled from eight experts' evidence — meaning the experts' own blind spots are baked into the model through the CoE annotations. When a new artifact type escapes all eight experts, both the data side and the reward side fail simultaneously, an inherent risk of the evidence-grounded paradigm.
The reward relies on Qwen-3-VL-8B as a judge. Reliability and bias analysis is deferred to Appendix F, but self-consistency bias remains a concern when the judge and the model under evaluation come from the same family.
On inference cost, REVEAL emits a full evidence chain, so its token count far exceeds that of a discriminative detector; the paper reports no latency or throughput comparison, which could be decisive in real-time review settings.
Promising directions: make the expert set extensible (adding an expert for a new artifact type while incrementally updating the CoE annotations), and add an explicit check to the reward for whether cited evidence genuinely comes from the corresponding expert output, further shrinking the room for fabricated evidence.

vs AIGI-Holmes: Both combine "small expert models + large-model reasoning", but AIGI-Holmes uses experts for global filtering and generates explanations afterwards through a multi-expert jury mechanism. REVEAL has experts emit structured artifact masks and diagnostic labels and writes that evidence explicitly into the CoE annotations, yielding a verifiable causal chain from low-level artifact to final verdict. The price is dependence on the eight experts' offline coverage.
vs RAIDX: RAIDX uses retrieval-augmented generation plus GRPO to improve textual descriptions, with a fairly generic reward design. REVEAL's rewards are forensics-specific and emphasize multi-view visual evidence alignment, which pins the explanation to the image itself rather than to retrieved priors.
vs low-level artifact detectors such as NPR / AIDE / HyperDet: They can even beat REVEAL in-domain (NPR reaches 95.40% on REVEAL-Bench) but collapse on new generators (59.37% average on REVEAL-Bench++). Low-level statistical cues and evidence-chain reasoning are therefore complementary: the former supply the evidence REVEAL depends on, the latter keeps judgments stable under distribution shift.
vs FakeBench / Holmes-Set: Those datasets offer image-text explanations, but the annotations are free-form and lack two dimensions — expert evidence and the reasoning process (see Table 1) — so they can only train a model that sounds right, not one that reasons correctly.

Rating¶

Novelty: ⭐⭐⭐⭐ Turning "the explanation must precede the verdict" into a probability factorization and optimizing the reasoning chain with three evidence-grounded rewards is a new angle for AIGC forensics, though the combination of expert evidence, CoT data, and GRPO is not itself unfamiliar.
Experimental Thoroughness: ⭐⭐⭐⭐ Covers in-domain, cross-domain GenImage, and a five-generator REVEAL-Bench++, with ablations on five backbones, training strategies, and rule-based aggregation, plus human preference and multi-view faithfulness studies; the deductions are the accuracy-only main metric and the unlabeled backbone in Table 5.
Writing Quality: ⭐⭐⭐ The motivation and the two-stage method read clearly, but the equations are badly mangled in typesetting (Equations (2), (3), (5), and (13) are all incomplete), the R-GRPO acronym is expanded inconsistently in two places, and the ablation table's metadata is incomplete.
Value: ⭐⭐⭐⭐ The 60K expert-evidence CoE dataset and a deployment form in which the eight experts never participate in inference are directly reusable for the explainable AIGC forensics line.