acceptodds
Under review as a conference paper at ICLR 2027

IVT-Guard: All-in-One Reasoning Model for AI-Generated Content Detection

Abstract

The rapid proliferation of highly realistic AI-Generated Content (AIGC) necessitates robust and interpretable detection mechanisms. However, existing detectors are predominantly confined to single modalities and provide binary outputs without reasoning. While Multimodal Large Language Models (MLLMs) present a promising solution, their development is constrained by the scarcity of multimodal reasoning data and the reasoning–detection optimization dilemma, where explicit reasoning supervision can compromise detection accuracy. To this end, we introduce IVT-Set, a comprehensive dataset comprising over 152K diverse image, video, and text samples equipped with multi-granularity Chain-of-Thought (CoT) reasoning trajectories. Based on it, we propose IVT-Guard, a pioneering framework for unified and interpretable AIGC detection across image, video, and text modalities. Furthermore, to overcome the aforementioned optimization dilemma, we design a novel three-stage training paradigm: Artifact-Aware Pre-training, Artifact-to-Evidence Supervised Fine-Tuning via artifact-aware injection, and Evidence-Verdict Consistency Group Relative Policy Optimization. Extensive experiments demonstrate that IVT-Guard achieves state-of-the-art detection performance across in-domain, out-of-domain, and cross-dataset settings while delivering faithful reasoning. Code and data will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-Generated Content Detection, Multimodal Large Language Models, Interpretable Reasoning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_6o48OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a large-scale multimodal AIGC detection dataset with multi-granularity CoT annotations and a unified detector for image, video, and text, motivated by the limitations of existing datasets and the lack of unified detection methods. To address the optimization dilemma between reasoning and detection, the paper proposes a three-stage training paradigm to improve artifact perception, evidence-grounded reasoning, and evidence-verdict consistency. Extensive experiments across various evaluation settings and benchmark datasets demonstrate strong detection performance and robust generalization.

Critical strengths
  1. The introduction of a large-scale multimodal AIGC detection dataset with multi-granularity CoT annotations provides an important resource. It addresses the limitations of existing datasets in terms of data scale, modality coverage, and annotation granularity.

  2. The paper is well motivated, and the proposed method is reasonably designed. The proposed three-stage training paradigm addresses the lengthy CoT optimization dilemma and shows generalizability across architectures. Extensive experiments and ablations support its effectiveness.

Critical weaknesses

Although Figure 2 compares different SFT strategies, the effect of response length on detection accuracy remains unclear. I suggest an additional analysis of different SFT response lengths to better support the paper's motivation.

Questions for authors
  1. The calculation of the OOD metric should be further clarified, particularly the inclusion of the "Real" column in the main tables but not in the ablation experiments.

  2. Figure 4 incorrectly uses "APP Loss" instead of "AAP Loss".

  3. The caption of Table 6 mentions three cross-dataset benchmarks, whereas four are reported.

  4. The quality threshold τ in Figure 3 is not explicitly specified, and the aggregation of evaluation scores could be clarified. The verification prompt and the content extracted by Reason(⋅) in Eq. (9) are not sufficiently specified.

4 comments

We thank the reviewer for their valuable feedback.

W1: Effect of SFT response length on detection accuracy.

We further investigate the effect of SFT response length by comparing the full response with four shortened variants. As shown in table A, the full response achieves the highest OOD accuracy across all three modalities. Shortening the reasoning components substantially degrades image and video accuracy by 6.56–8.35 and 21.77–23.97 percentage points, respectively. These results indicate that simply shortening CoT is not an effective way to alleviate the reasoning–detection optimization dilemma.

Table A: OOD accuracy (%) and average SFT response length (tokens, in parentheses).

SFT response variantImage OODVideo OODText OOD
Full response (Ours)92.27 (577)78.40 (568)97.63 (681)
Short thinking84.00 (357)56.20 (348)96.58 (397)
Short explanation85.71 (486)54.43 (482)96.21 (536)
Short thinking + explanation83.92 (265)56.63 (262)97.42 (253)
Short CoT (thinking + answer)85.35 (333)54.88 (329)96.54 (387)

Q1: Clarification of OOD metric calculation.

For each modality, each unseen generator is evaluated on 400 fake samples together with the same shared set of 400 real samples. Thus, each generator column in the main tables reports accuracy over these 800 samples, while the “Real” column separately reports accuracy on the 400 real samples.

In the ablation tables, OOD accuracy is the macro-average across unseen generators, with equal weight for each generator. The “Real” column is excluded from this average because real-sample performance is already included in each generator's accuracy.

Q2 & Q3: Corrections to figure labels and table captions.

We thank the reviewer for pointing out these errors. We make the following corrections:

(1) Figure 4: Corrected the label from "APP Loss" to "AAP Loss".

(2) Table 6: Corrected the caption from "three cross-dataset benchmarks" to "four cross-dataset benchmarks" to match the four benchmarks reported in the table.

Q4: Clarification of evaluation thresholds, score aggregation, and verification procedure.

We clarify the following details:

(1) Quality thresholds and score aggregation. Each judge scores six dimensions (1–5), with scores averaged across dimensions and judges. The review and acceptance thresholds are 3.0 and 4.5, respectively. Samples below the review threshold undergo human revision. Additionally, acceptance requires every judge to score at least 4.5 on each dimension. Otherwise, iterative refinement continues for up to three rounds.

(2) Verification prompt and reasoning extraction. We clarify that Reason⁡(⋅)\operatorname{Reason}(\cdot) extracts and concatenates the <thinking>, <key_features>, and <explanation> fields, excluding <answer>. We add the verification prompt for Qwen3-4B-Instruct-2507 in Appendix fig. 24.

Reviewer_6o48OpenReview ↗

Response to the rebuttal

The authors have addressed my questions. Given the contributions of the dataset and training paradigm to the field of AIGC detection, I will keep my recommendation to accept the paper.

We thank the reviewer for their positive feedback. We will incorporate these revisions into the final version.

Thank you for recommending acceptance. We would appreciate it if you could update your rating.

Reviewer_GG6dOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes IVT-Guard, a unified detection framework for AI-generated image, video, and text content, and constructs IVT-Set, a dataset containing approximately 152K samples with multi-level CoT annotations. The method adopts a three-stage training strategy: first, Artifact-Aware Pre-training learns forgery-detection representations for images and videos; then, Artifact-Aware Injection fuses these representations into an MLLM with CoT-supervised fine-tuning; finally, Evidence-Verdict Consistency GRPO reinforces the consistency between reasoning and the final judgment. Experiments demonstrate strong performance on ID, OOD, and cross-dataset detection, and provide an evaluation of reasoning-explanation quality.

Critical strengths
  1. A unified detection framework and a relatively comprehensive dataset. This paper attempts to integrate image, video, and text forgery detection into a unified MLLM framework and constructs a multimodal dataset covering many generative models, which has potential research and application value.
  2. Strong empirical performance. The method achieves clear improvements in image and video OOD detection, and the paper also provides ablation experiments on training stages, feature fusion, model backbones, and robustness.
Critical weaknesses
  1. Reasoning faithfulness is insufficiently validated. The CoT annotations were generated with knowledge of the ground-truth labels, while the judge in EVC-GRPO only checks consistency between the explanation text and the final conclusion and cannot verify whether the evidence actually exists. Therefore, the current experiments are not yet sufficient to support the paper’s core claim of faithful/evidence-grounded reasoning.
  2. The source of the main performance gains remains unclear. The proposed AAP and A2E-SFT introduce additional forensic encoders and feature fusion modules. However, the paper lacks a comparison between Answer-Only and CoT supervision under the same artifact-enhanced architecture. Therefore, it remains unclear whether the gains arise from alleviating the reasoning–detection dilemma or primarily from stronger forensic representations.
Questions for authors

Q1. EVC-GRPO’s gain over Vanilla GRPO is limited. Can the authors verify its contribution to genuine evidence faithfulness, not just text-level consistency via independent evidence-correctness or hallucination-rate evaluation?

Q2. Video TPA uses cross-frame average pooling, and frame shuffling barely affects accuracy. How much does the model truly exploit temporal artifacts rather than single-frame spatial features?

Q3. Given existing cross-modal AIGC detection frameworks, the unified design mainly integrates modality-specific encoders, feature injection, and existing SFT/GRPO. Can the authors clarify the core methodological novelty beyond broader modality coverage?

Q4. The framework adds a visual encoder and generates long CoTs. Can the authors compare latency, memory, and throughput against conventional discriminative detectors to assess practical cost?

If these key questions are convincingly addressed with supplementary evidence, I am willing to reconsider and raise the score.

3 comments

We thank the reviewer for their valuable feedback.

W1 & Q1: Evaluation of reasoning faithfulness.

We use Qwen3-32B to decompose the reasoning steps and explanations into atomic claims and define two metrics: Evidence Support Rate (ESR), measuring the proportion of claims supported by the input, and Hallucination Rate (HR), measuring the proportion of responses containing at least one contradicted claim. Two annotators independently inspect the corresponding original modality inputs (images, videos, or texts) and classify each atomic claim as supported (S), contradicted (C), or unverifiable (U), with disagreements resolved by a third annotator. Both metrics are calculated using the adjudicated annotations on the OOD test set. Specifically, ESR is calculated as S/(S+C+U)S/(S+C+U), while HR is the proportion of responses containing at least one contradicted claim.

As shown in Table A, compared with GRPO, EVC-GRPO increases ESR by 5.06, 2.96, and 2.49 pp and reduces HR by 3.39, 7.83, and 8.25 pp for image, video, and text, respectively. These results demonstrate that EVC-GRPO consistently improves evidence faithfulness and reduces hallucinations across all three modalities.

Table A: Comparison of reasoning faithfulness between GRPO and EVC-GRPO across modalities under the OOD setting.

MethodModalityclaimsSCUESR (%)HR (%)
GRPOImage42,91336,9952,3103,60886.2110.64
Video37,23131,3932,6293,20984.3218.46
Text23,59120,6021,0521,93787.3316.38
EVC-GRPOImage50,23945,8532,5191,86791.277.25
Video43,91238,3262,5643,02287.2810.63
Text23,07320,7248471,50289.828.13

W2: Comparison between Answer-Only and CoT supervision under matched training configurations

As shown in table B, we compare Answer-Only and CoT supervision across three training configurations: standard SFT, AAP + A2E-SFT, and AAP + A2E-SFT + GRPO. Compared with standard SFT, AAP + A2E-SFT yields substantially larger OOD gains under CoT than Answer-Only supervision for image (+6.31 vs. −0.88 pp) and video (+24.22 vs. +4.40 pp). GRPO further improves CoT-based video OOD accuracy to 81.88%, while maintaining strong performance across all three modalities. These results suggest that while strong artifact-aware representations are essential, combining them with CoT supervision to alleviate the reasoning–detection optimization dilemma can substantially improve detection performance.

Table B: Comparison of Answer-Only and CoT supervision across training configurations under the OOD setting.

SupervisionTrainingImageVideoText
Answer-OnlySFT90.4874.5899.54
AAP + A2E-SFT89.6078.9899.50
AAP + A2E-SFT + GRPO89.4879.9899.38
CoTSFT85.9654.1897.46
AAP + A2E-SFT92.2778.4097.63
AAP + A2E-SFT + GRPO92.0081.8897.87

Q2: Contribution of temporal artifacts to video detection.

We acknowledge that "temporal artifacts" is not an accurate description of TPA and will revise it to "frame-level artifacts". Frame shuffling has little impact on detection accuracy, suggesting that relying on frame-level cues enables our model to maintain robust detection under video input perturbations.

To evaluate the contribution of multi-frame information, we replace PE's multi-frame input with a randomly sampled frame repeated to match the original frame count during both training and inference, while keeping Qwen's multi-frame input unchanged. As shown in Table C, distinct multi-frame inputs improve accuracy in four of five evaluation settings, with an 8.75 pp gain on OOD, demonstrating the benefit of aggregating frame-level artifacts across different frames.

Table C: Comparison of multi-frame and repeated single-frame PE inputs on video detection accuracy (%).

VariantIDAIGVDBenchGenBuster++LOKIOOD
Repeated single-frame88.5088.8974.7087.9674.05
Original multi-frame92.0088.6477.6089.2582.80

Q3: Clarification of methodological novelty.

The core methodological novelty of IVT-Guard lies in a three-stage training paradigm for multimodal AIGC detection that progressively connects forensic artifact representation learning, evidence-grounded reasoning, and consistency-guided policy optimization, thereby alleviating the optimization dilemma between reasoning and detection. Specifically, (1) AAP employs binary authenticity supervision to explicitly train visual encoders to capture forensic artifact representations; (2) A2E-SFT introduces complementary global-local injection for images and temporal-spatial injection for videos, integrating forensic representations into the MLLM's visual stream and using CoT supervision to guide evidence-grounded reasoning; and (3) EVC-GRPO introduces a dedicated evidence-verdict consistency reward that evaluates the agreement among the reasoning-derived verdict, the model's final prediction, and the ground-truth label, explicitly penalizing inconsistencies during policy optimization.

Q4: Comparison of computational efficiency with conventional discriminative detectors.

We compare end-to-end inference latency, peak GPU memory, and throughput on a single NVIDIA H800 using 500 OOD samples per modality. All methods use batch size 1 for latency and memory and batch size 4 for throughput. OOD accuracy is averaged across generators.

Although IVT-Guard incurs higher computational costs, it provides explicit forensic reasoning and evidence-grounded explanations. We will also release IVT-Guard-fast, which directly outputs detection results without producing an explicit CoT, reducing inference latency and peak GPU memory usage while increasing throughput, and achieves the highest OOD accuracy among conventional detectors across all three modalities.

Table D: OOD accuracy and efficiency of image detectors.

MethodOOD Acc. (%)Latency (s)Peak GPU Memory (GB)Throughput (samples/s)
UnivFD76.060.01812.54870.059
DIRE73.861.17692.8841.517
NPR65.790.00880.786131.603
AIDE80.480.08404.58412.752
DGS-Net83.480.04794.27223.540
PGC80.920.02022.01779.583
IVT-Guard-fast89.480.092514.19115.183
IVT-Guard93.482.887614.8041.160

Table E: OOD accuracy and efficiency of video detectors.

MethodOOD Acc. (%)Latency (s)Memory (GB)Throughput (samples/s)
ReStraV55.970.35861.0212.725
NSG-VD47.450.50319.1902.007
DeMamba67.430.45822.3452.207
WaveRep72.250.62115.6161.621
IVT-Guard-fast79.980.339116.2843.297
IVT-Guard82.803.041016.9510.893

Table F: OOD accuracy and efficiency of text detectors.

MethodOOD Acc. (%)Latency (s)Memory (GB)Throughput (samples/s)
ModernBERT (public)64.420.01571.464247.430
Desklib AI92.330.01722.69187.436
DivEye64.670.01182.31198.983
FourierGPT51.250.09429.39514.612
DetectAnyLLM-GPT-Neo-2.7B87.370.021710.43869.456
WaveDetect95.250.01562.500117.171
IVT-Guard-fast99.380.044911.62067.675
IVT-Guard98.003.269912.7951.071

Dear reviewer GG6d, this is a gentle reminder regarding our rebuttal. We would be very grateful if you could take a moment to review our responses when you have time. Thank you again for your time and effort.

Reviewer_cZTzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces IVT-Set, a dataset of about 152K image, video, and text samples with structured explanations, and IVT-Guard, a unified detector based on Qwen3-VL-8B. Training combines artifact-aware visual encoder pre-training, supervised fine-tuning with injected forensic features, and GRPO with an explanation–verdict consistency reward. The experiments cover in-domain, unseen-generator, and cross-dataset detection, with ablations of training stages, feature injection, and modality combinations.

Critical strengths
  1. The dataset and unified system offer practical value. IVT-Set combines three modalities and 44 generators with structured explanation supervision. The construction makes a meaningful effort to reduce easy shortcuts by matching image content and quality, standardizing video encoding, and controlling text length and domain. These choices make the resource more useful than simply combining existing datasets.

  2. The empirical support goes beyond a single benchmark or generic model comparison. The paper compares several specialized detectors and evaluates transfer to external datasets and unseen generators. The ablations support the usefulness of the added perception modules, especially for video: Table 4 reports an increase in video OOD accuracy from 54.18% with standard SFT to 78.40% with AAP and A2E-SFT. I view the dataset, working system, and these results as meaningful contributions even though some explanation claims need stronger evidence.

Critical weaknesses
  1. Human validation does not yet establish whether the explanations describe real evidence. Section 3.2 gives the annotation model the correct label, and Appendix A.4 asks it to provide features supporting that label. This can produce plausible reasons even when the stated cues are absent. The paper describes human revision, but does not provide enough quantitative detail to assess the remaining error rate. The EVC judge reads the generated reasoning rather than the original visual input, so agreement with the verdict does not independently verify the evidence. Please audit a random subset stratified by modality, label, and generator, including annotations that passed automatic review. Have at least two independent annotators inspect the original input and assess each claimed cue without seeing the supplied label or model identity. Report sample counts, the proportion of unsupported claims, annotator agreement, and examples of corrections. Also report the existing human-review coverage and acceptance thresholds, and compare automatic quality scores with human judgments. Keep claims about faithful reasoning proportionate to these results.

  2. Important comparisons are missing, and training conditions need clearer alignment. Key baselines. Table 7 compares explanation quality only with generic MLLMs. Please add specialized explainable detectors, prioritizing AIGI-Holmes for images and VideoVeritas or VidGuard-R1 for videos, under the same evaluation rubric and blinded human assessment. The detection tables would also benefit from Effort and WaveRep (Seeing What Matters). These works were first public before June 26, 2026. Please prioritize direct competitors and explain any practical barriers to evaluating them; an exhaustive list is unnecessary. Fair comparisons. Section 5.1 states that modality-specific detectors are retrained on IVT-Set, but does not make equally clear whether specialized MLLM baselines receive comparable training data. Please list each baseline’s checkpoint, training data, adaptation steps, and inference settings. Include a controlled comparison with a closely related method using the same backbone and IVT-Set subset, with comparable training and inference budgets where possible. Report unavoidable differences. This would help separate improvements due to data adaptation and added encoders from those due to the proposed training strategy. The reward comparison also needs uncertainty estimates. In Table 4, EVC-GRPO improves OOD accuracy over ordinary GRPO by 1.48, 0.92, and 0.13 percentage points for image, video, and text. Please repeat this comparison across seeds and report variability, particularly for the small text gain.

Questions for authors
  1. What proportion of annotations received human review, and was a random sample of automatically accepted annotations independently checked for unsupported evidence?

  2. Which specialized MLLM baselines were adapted to IVT-Set, and did they use the same training examples and supervision as IVT-Guard?

  3. Are splits grouped by the original image, video, or document so that all derived samples stay together? Please describe deduplication against external test sets and clarify the AAP training split.

6 comments

We thank the reviewer for their valuable feedback.

W1: Independent human validation of evidence faithfulness.

We conduct an independent human audit of IVT-Set. IVT-Set contains 152,196 annotations. We define sampling strata by modality × real/fake label × source, yielding 25 image, 16 video, and 16 text strata. We randomly sample up to 200 records per stratum, resulting in 11,176 annotations (4,776 images, 3,200 videos, and 3,200 texts), containing 58,460 cues listed in the key_features field.

Two annotators independently inspect the original inputs and assess each cue for evidential support. Both annotators are blinded to the ground-truth labels and model identities, and disagreements are resolved by a third annotator through independent adjudication. The audit identifies 2,277 cues as unsupported (3.89% overall; 3.24% for images, 4.05% for videos, and 4.70% for texts), with 94.16% (55,044/58,460) raw agreement between annotators before adjudication.

We add representative examples of unsupported cues and their human corrections in figures 25–30 of the Appendix.

Q1: Human review coverage and independent verification of automatically accepted annotations.

In the original annotation pipeline, 37,574 annotations (24.69%) received human review. The review and acceptance thresholds were 3.0 and 4.5, respectively, on the automatic quality scale.

Of the 11,176 audited annotations, 8,417 were accepted through the automatic critique–refinement pipeline and 2,759 were routed for human review, containing 38,137 and 20,323 cues, respectively. Among automatically accepted annotations, 1,991/38,137 cues (5.22%) are unsupported. In contrast, following human revision, only 286/20,323 cues (1.41%) remain unsupported.

W2 & Q2: Missing key baselines, fair training comparisons, and statistical reliability.

Key baselines.

(1) Explanation quality. We revise table 8 (formerly table 7) in the Appendix to add specialized explainable detectors, including UniGenDet, Veritas, FakeVLM, and AIGI-Holmes for images, and Skyra-RL and VideoVeritas for videos. As the Visual Expert weights for AIGI-Holmes are not yet available, we reproduced AIGI-Holmes without the DPO stage.

Since Elo ratings depend on the comparison pool, we recomputed all ratings after adding these baselines. The values in parentheses denote (explanation score evaluated by Qwen3.5-122B-A10B, human-evaluated Elo rating).

  • Image: UniGenDet (3.85, 677.1), Veritas (3.49, 748.5), FakeVLM (3.01, 650.0), and AIGI-Holmes (4.46, 1270.8), compared with IVT-Guard (4.70, 1324.1).
  • Video: Skyra-RL (2.45, 703.0) and VideoVeritas (3.16, 1234.5), compared with IVT-Guard (4.34, 1250.0).

IVT-Guard achieves higher explanation scores and human Elo ratings than these specialized baselines in both modalities.

(2) Detection performance. We revise tables 2 and 3 in the main text to add Effort for image detection and WaveRep (Seeing What Matters) for video detection, respectively, with both additions highlighted in blue. We retrained both baselines on IVT-Set following their respective training configurations. Our method achieves superior overall detection performance compared with both Effort and WaveRep.

Table A: Image detection accuracy (%) of Effort and IVT-Guard. OOD denotes the average accuracy on Real, FLUX.1 Krea, Qwen-Image, Midjourney v6, Infinity, Janus-Pro-7B, and LlamaGen.

MethodIDChameleonLOKIGenImageGenBuster++OOD (with Real)Avg.
Effort98.0059.6286.3486.3352.0589.8986.33
IVT-Guard (Ours)98.7593.3287.0293.3766.3094.2492.66

Table B: Video detection accuracy (%) of WaveRep and IVT-Guard. OOD denotes the average accuracy on Real, EasyAnimateV5.1, HunyuanVideo-I2V, Kling 2.1, Veo 3, and Vidu Q2.

MethodIDLOKIAIGVDBench*GenBuster++OOD (with Real)Avg.
WaveRep91.6374.8062.8766.8576.1778.66
IVT-Guard (Ours)92.0089.2588.6477.6083.2986.82

Fair comparisons.

(1) Implementation details of specialized MLLM baselines.

Following the evaluation protocols adopted in prior work (e.g., UniGenDet, Skyra, and VideoVeritas), we evaluate specialized MLLM baselines using their respective released checkpoints and original evaluation protocols, including prompts and inference configurations. We add a new table (table 6) to the Appendix, summarizing the checkpoints, training data, adaptation steps, and inference settings of each baseline.

(2) Controlled comparison with Skyra.

As shown in table C, we retrain Skyra using the same Qwen3-VL-8B-Instruct backbone and IVT-Set training split as IVT-Guard, with comparable training and inference budgets, and further retrain a variant augmented with pretrained CLIP and PE encoders. The unavoidable differences lie in the training strategy: Skyra follows its original SFT-to-GRPO pipeline, whereas IVT-Guard adopts AAP, A2E-SFT, and EVC-GRPO. These results suggest that our proposed three-stage training strategy is a key driver of the performance gains.

Table C: Controlled comparison of video detection accuracy (%).

MethodIDLOKIAIGVDBench*GenBuster++OODAvg.
Skyra77.5062.2861.9565.0565.6368.74
Skyra (with CLIP and PE)70.7563.7262.7265.8558.9364.59
IVT-Guard (Ours)92.0089.2588.6477.6082.8086.65

Report unavoidable differences.

(1) Multi-seed comparison of GRPO and EVC-GRPO. We repeat the comparison using three random seeds (42, 44, and 46) and report the mean ± sample standard deviation for both methods in table D. EVC-GRPO improves OOD accuracy by 2.37, 1.06, and 0.86 percentage points on image, video, and text, respectively. These results demonstrate consistent OOD improvements across all three modalities.

Table D: ID and OOD accuracy (%) of GRPO and EVC-GRPO across three seeds.

MethodImage IDImage OODVideo IDVideo OODText IDText OOD
GRPO98.04 ± 0.2691.57 ± 0.4892.17 ± 0.3681.32 ± 0.1398.50 ± 0.3397.52 ± 0.24
EVC-GRPO98.38 ± 0.3393.94 ± 0.6492.67 ± 0.8882.38 ± 0.5998.50 ± 0.2298.38 ± 0.27

Q3: Clarification of data splits, external test set deduplication, and AAP training split.

Data split. We group each original sample and its derived samples by source ID. The training set primarily contains two types of samples: (1) pairs of real and fake samples sharing the same source ID, which provide controlled supervision for learning authenticity-related forensic cues under similar semantic content; and (2) unpaired real or fake samples, which increase source diversity and reduce reliance on explicitly matched pairs.

Deduplication against external test sets. We audited all 45,000 training samples against the 11 external test sets. For images and videos, SHA-256 was used for exact matching, while perceptual hashing and DINOv2 similarity were used to detect visual near-duplicates. For text, we used normalized exact matching, MinHash for lexical overlap, and MiniLM for semantic similarity. We further examined all flagged pairs and excluded a total of 13 duplicate samples from the external test sets.

AAP training split. AAP is trained on a total of 45,000 in-distribution (ID) examples, including 30,000 from the A2E-SFT split (10,000 per modality) and 15,000 from the EVC-GRPO split (5,000 per modality), with balanced real/fake labels in each modality. The two splits use different examples, and out-of-distribution (OOD) data are used exclusively for testing.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.