acceptodds
Under review as a conference paper at ICLR 2027

CLOVER: Coverage-Led Optimization of Verifiable Evidence Reasoning for AI-Generated Image Detection

Abstract

AI-generated image detectors must recognize forgeries across diverse generators and scenes and provide explanations. Binary labels are inexpensive to obtain but provide no supervision for explanations, whereas evidence annotations are costly and may still omit valid observations. To this end, we introduce CLOVER, a training framework that expands detection coverage through binary supervision and advances explanation learning through open-ended evidence verification. For coverage-led supervised fine-tuning, we combine a large binary-labeled image collection for detection with a smaller, scene-diverse evidence corpus for explanation learning. This allows new forgery sources to enter training at low annotation cost with only binary labels. Atomic evidence reinforcement learning then improves evidence quality in two stages: first encouraging the discovery of more visually supported observations, then strengthening penalties for unsupported or uncertain claims. During this process, we employ a self-judge to assess the model's open-ended evidence against the image one claim at a time, enabling rewards that require authenticity labels but no reference explanations. Experiments show that supervised training already achieves strong detection and explanation performance with an evidence corpus roughly 10% the size of the binary corpus. Subsequent evidence optimization further improves precision and reduces hallucinations, demonstrating that visual verification can improve explanation reliability without collecting additional reference explanations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, vision-language models, evidence reasoning, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1vSoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CLOVER, a training framework for explainable AI-generated image detection. It pairs a large binary-labeled corpus, which provides broad coverage of forgery sources, with a much smaller scene-diverse evidence corpus. These are used for joint supervised fine-tuning on binary detection, explainable detection, and label-conditioned explanation. Two-stage reinforcement learning then follows, using only authenticity labels. A frozen self-judge verifies each evidence claim against the image, and a separate judge groups overlapping claims so that repeated evidence is rewarded once. The first stage rewards discovering supported evidence, and the second strengthens penalties for uncertain or invalid claims. Across five detection and three explanation benchmarks, with ablations and a blinded expert assessment, the SFT model achieves the best average detection macro-F1 among the compared methods. RL further improves explanation precision and reduces hallucinations with little change in detection.

Critical strengths
  • Explanation improvements without reference explanations. The claim-level self-judge and semantic grouping reduce hallucinations and improve precision.
Critical weaknesses
  • Generator overlap complicates the attribution of detection gains. CLOVER is trained on 35 generator groups, and Table 7 shows that several evaluation sets share generator families with training. Evaluation on fully held-out generator families, together with representative baselines retrained under matched generator coverage, would help separate methodological gains from the benefits of broader training data.
  • The evaluation provides limited evidence of reliability under realistic online conditions. The paper reports no robustness tests against common distribution shifts such as JPEG compression, resizing, and reposting, and its only in-the-wild set contains no real images, so false positives remain untested. Robustness evaluation under these perturbations, together with results on one or two additional in-the-wild benchmarks containing real images (e.g., Chameleon, WildRF, SocialRF, or SynthWildX), would provide stronger evidence of practical reliability.
  • The general-purpose MLLM comparisons could be more current. The strongest general-purpose baselines come from the Qwen3-VL and InternVL3.5 families, while more recent open-weight multimodal models such as Qwen3.6-27B are not evaluated. Evaluating such models under the same protocols would clarify whether CLOVER's explanation advantages persist.
Questions for authors

Please address the points raised in the Weaknesses section

2 comments
AuthorsOpenReview ↗

Response to Reviewer 1vSo (1/2): Generator Generalization and Matched-Coverage Evaluation

W1. Generator overlap and attribution of detection gains.

Unseen-generator tests and matched-coverage retraining both support CLOVER’s detection advantage. Controlled ablations also validate specific training choices.

(a) CLOVER retains its lead on entirely unseen generators. Z-Image, HunyuanImage-3.0, and BAGEL were excluded from task training and model selection. Using the original checkpoint, CLOVER-RL achieves 79.90 average macro-F1, exceeding Qwen3-VL-8B and FakeVLM by 18.70 and 10.00 points, respectively, and leading on every family.

After excluding the Nano Banana and GPT Image families from all training stages and retraining with matched data volume and training budgets, CLOVER-RL still achieves 76.10 macro-F1 on Nano-banana, exceeding FakeVLM by 7.59 points, and 84.00% fake recall on GPT-Image2.

(b) With matched training coverage, CLOVER leads on all five benchmarks. We retrain D³, the strongest binary-only baseline in our evaluation, on the same 100K binary-labeled images, matching generator sources and the real/fake ratio.

MethodGenImageLOKI-DetSoFake-OODFakeBenchAIGIBenchAverage
D³97.9174.2272.7286.0178.2381.82
D³, matched coverage97.4679.6886.9292.1588.6488.97
CLOVER-SFT98.1183.9493.7296.7289.7192.44
CLOVER-RL97.9383.0093.7096.1091.0192.35

Detection macro-F1 (%), averaged equally across five benchmarks.

Broader coverage improves D³ by 7.15 points, but CLOVER-SFT/RL still lead by 3.47/3.38 points. Training coverage contributes to performance; matching it does not eliminate CLOVER’s advantage.

(c) Same-data, same-backbone evidence supports the label-conditioned task. With data, backbone, and total training presentations fixed, Table 14(b) shows gains of 3.44 points in explanation precision and 2.96 points in TriDF-B F₀.₅, while maintaining detection performance. This directly validates the training design.

AuthorsOpenReview ↗

Response to Reviewer 1vSo (2/2): Real-World Reliability and Updated MLLM Comparisons

W2. Reliability under realistic online conditions.

CLOVER has lower false-positive rates on real in-the-wild images than the compared baselines. Additional experiments quantify corruption effects on both detection and explanation.

(a) The original evaluation includes real social-media images. SoFake-OOD-v3 includes platform-native real images and preserves their original dimensions and compression characteristics. CLOVER-RL achieves 93.70% macro-F1 on this benchmark, whose evaluation content is disjoint from training. Official dataset description.

(b) Independent in-the-wild real images directly test false positives. We evaluate all models on the same 1,000 independent real in-the-wild images, with no training-content overlap and fixed decision rules.

MethodGPT-Image2-wild fake recall ↑In-the-wild real-image FPR ↓
Qwen3-VL-8B77.4018.20
FakeVLM59.1324.60
CLOVER-SFT86.739.50
CLOVER-RL85.339.70

Values in %. Recall and FPR are measured on separate fake-image and real-image sets.

Relative to Qwen3-VL-8B, CLOVER-RL increases fake recall by 7.93 points and reduces real-image FPR by 8.50 points, improving both detection and false-positive control.

(c) Corruption effects are quantified across five detection and three explanation benchmarks. All conditions use the original CLOVER-RL checkpoint and fixed inference settings, without adaptation or threshold tuning.

ConditionDetection avg. macro-F1 ↑FC Precision ↑FC Hall. Rate ↓FC Evi@Img ↑FC Notice rate
Original92.3561.469.2970.4317.89
JPEG Q7091.0660.109.8068.8019.30
JPEG Q3087.0655.2011.6063.1026.40
Resize 0.5×90.0458.4010.2066.9022.20
Gaussian blur (σ=1.5)88.3056.7010.9064.8024.30
Screenshot89.5057.8010.6065.9023.20

Detection is averaged equally across five benchmarks. FC denotes FakeClue-Evidence; Notices and empty explanations remain in its evaluation denominator.

JPEG Q70, resizing, and screenshots reduce average detection macro-F1 by 1.29, 2.31, and 2.85 points, respectively. JPEG Q30 reduces it by 5.29 points, while supported-image coverage decreases by 7.33 points, Notice rate rises by 8.51 points, and hallucination rate rises by 2.31 points. Under severe corruption, the model abstains more often, but additional evidence loss and unreliable claims remain.

Screenshots use a fixed browser, 100% page zoom, DPR=1, and aspect-preserving display at a 512-px long edge, capturing only the image region. TriDF measures retention of the original reference evidence; LOKI uses corrupted full images and corresponding crops extracted after corruption; FC verifies claims against the corrupted images.

W3. More recent general-purpose MLLM baselines.

With Qwen3.6-27B added, CLOVER-RL still leads in detection and every reported explanation metric.

MethodDetection avg. macro-F1 ↑TriDF-A F₀.₅ ↑TriDF-B F₀.₅ ↑LOKI Overall ↑FC Precision ↑FC Hall. Rate ↓FC Evi@Img ↑
Qwen3.6-27B81.4138.4126.8271.2633.6010.9640.47
CLOVER-SFT92.4443.4736.1567.6954.8414.5565.48
CLOVER-RL92.3548.6241.0173.5261.469.2970.43

FC denotes FakeClue-Evidence.

CLOVER-RL leads by 10.94 points in average detection macro-F1, 27.86 points in FC precision, and 29.96 points in supported-image coverage, while reducing the hallucination rate by 1.67 points. Its explanation advantage spans precision, coverage, and hallucination control.

Reviewer_oFuMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

CLOVER trains an explainable AI-generated image detector on Qwen3-VL-8B. It separates supervision by purpose: 100K binary images from 35 generator groups give source coverage for detection, and 11K evidence-annotated images from 7 scene domains, written by GPT-5.5 and Claude Opus 4.8, teach explanation. Both are trained jointly with a label-conditioned explanation task. RL then refines the explanations. A frozen copy of the SFT model labels each claim as supported, uncertain, or hallucinated, a Qwen3-32B judge merges overlapping claims, and two GSPO stages apply a coverage reward and then a faithfulness reward, using only authenticity labels. The paper reports 92.4 average detection macro-F1 across five benchmarks. RL improves explanation metrics, for example FakeClue-Evidence precision rises from 54.8 to 61.5 and the hallucination rate falls from 14.6 to 9.3.

Critical strengths
  1. Splitting supervision so that binary labels give source coverage and evidence labels give scene coverage is sensible and validated: going from 2 to 7 scenes raises explanation precision from 47.6 to 54.8, while generator diversity drives detection.

  1. The RL reward needs only authenticity labels and works at the level of semantic groups, so each observation is credited once and hallucinated paraphrases cannot hide inside supported groups.

  1. Extra-SFT, no-claim-judge, reward-order, grouping, and abstention ablations show that the RL gains over CLOVER-SFT.
Critical weaknesses
  1. Every detection benchmark overlaps the training generator families: AIGIBench is a split of a training corpus, and Nano Banana and GPT Image 2 are among the training groups. There is no way to know whether the performance is due to the dataset or the method.

  1. Several baselines are misconfigured or out of scope. Effort scores 76.6% on GenImage here versus 91.1% in its own paper. UniFD is about 14 points below published results. ProDet and VERITAS are face-deepfake detectors evaluated on general-scene benchmarks. No threshold-free metrics such as AUROC are reported.

  1. CLOVER’s own binary-only model (93.1) outperforms every other variant. With no same-data or same-backbone baselines, the corpus and backbone can explain both the detection lead and the explanation lead over prior methods; only the RL-over-SFT gain is attributable to the method.

  1. Claude Opus 4.8 writes and reviews its own annotations. Both the self-judge and the Actor come from that SFT model, so the RL rewards self-consistency. The self-judge outperforming the 32B and 38B judges is consistent with shared bias, and expert agreement is reported only for the self-judge.

  1. The decision is generated before the explanation, so explanations are post-hoc. The reward measures visual grounding; the explanations do not causally verify the decision-making; there is no causal test, and label-conditioned training may teach the model to rationalize any label it is given.

  1. The annotation teachers (GPT-5.5, Claude Opus 4.8), FakeXplainer, Defake-o3 and PAVE are not compared.
Questions for authors

Please see the weakness section.

8 comments
AuthorsOpenReview ↗

Response to Reviewer oFuM (1/3): Generalization and Baseline Evaluation

W1. Generator overlap and generalization.

The original paper already includes cross-dataset OOD evidence. Additional experiments show that CLOVER retains its detection advantage without exposure to the target generators.

(a) The original evaluation includes OOD benchmarks. SoFake-OOD and FakeBench are strictly disjoint from the training data at both the image and content-group levels and include some generation sources unseen during task training (Table 7). Overlap in generator families does not imply an absence of out-of-distribution evaluation.

(b) CLOVER consistently leads on three entirely unseen generator families. Z-Image, HunyuanImage-3.0, and BAGEL were excluded from all task-training and model-selection stages. We evaluate the original checkpoints on identical test images, without adaptation or threshold tuning.

MethodZ-ImageHunyuanImage-3.0BAGELAverage
Qwen3-VL-8B60.8058.5064.3061.20
FakeVLM69.9067.4072.4069.90
CLOVER-SFT80.3078.8081.2080.10
CLOVER-RL80.1078.6081.0079.90

Macro-F1 (%), averaged equally across generators.

CLOVER-RL exceeds Qwen3-VL-8B and FakeVLM by 18.70 and 10.00 points, respectively, on average.

(c) CLOVER retains its advantage after removing the target generator families. We exclude the Nano Banana and GPT Image families from all task-training stages and retrain with matched data volume and training budgets. CLOVER-RL still achieves 76.10 macro-F1 on Nano-banana, exceeding FakeVLM by 7.59 points. On GPT-Image2, its fake-image recall reaches 84.00%, exceeding Qwen3-VL-8B by 6.60 points, while the false-positive rate on an independent real-image set decreases from 18.20% to 10.40%.

W2. Baseline configuration, applicability, and evaluation metrics.

Our baselines follow their papers and official implementations. CLOVER’s lead does not depend on low scores from face-forgery detectors, and additional AUROC results support its advantage.

(a) Score differences across evaluation conditions do not establish a configuration error. We have checked each baseline’s configuration and followed its official inference settings. Effort and UniFD achieve GenImage accuracy of 76.60% and 65.34% in our evaluation, compared with 91.10% and 79.50% in the cited literature; the UniFD figure comes from a subsequent publication. We reevaluate all methods on a common set of selected test images, under conditions that differ from those of the cited results.

Those published scores are also below CLOVER-SFT/RL’s 98.14%/97.93%.

(b) CLOVER’s lead does not depend on ProDet or VERITAS. CLOVER-RL achieves 92.35% average macro-F1 across five benchmarks, exceeding the strongest external baseline, FakeVLM (86.31%), by 6.04 points. We will move ProDet to the supplementary material as a cross-domain transfer reference. The VERITAS paper also reports results on general AIGC benchmarks, so categorically treating it as inapplicable is inaccurate. We retain it as a baseline and explicitly note its face-domain training scope.

(c) Threshold-free evaluation also supports CLOVER’s advantage. We compute AUROC from continuous authenticity scores using the same models and test images, without tuning thresholds on test labels.

MethodAverage macro-F1Average AUROC
FakeVLM86.3193.18
CLOVER-SFT92.4497.21
CLOVER-RL92.3597.16

Both metrics are averaged equally across the five benchmarks.

CLOVER-RL exceeds FakeVLM by 3.98 AUROC points, consistent with its macro-F1 advantage.

AuthorsOpenReview ↗

Response to Reviewer oFuM (2/3): Methodological Contributions and Self-Judge

W3. Detection–explanation trade-off and methodological contribution.

The original controlled experiments separately validate supervision coverage, the label-conditioned task, and RL. The claim that “only the RL-over-SFT gain is attributable to the method” overlooks this evidence.

(a) The label-conditioned task directly improves explanations with the same data and backbone. Table 14(b) controls the data, backbone, and total number of training presentations. Explanation precision improves by 3.44 points and TriDF-B F₀.₅ by 2.96 points, while detection performance is maintained. This is precisely a same-data, same-backbone comparison.

(b) Fixed-budget ablations validate the supervision-coverage design. Table 3(b) shows that broader generator-source coverage improves detection. Table 14(a) shows that expanding scene coverage from two to seven domains increases explanation precision from 47.64 to 54.84. How supervision is organized and allocated is itself a methodological contribution.

(c) Joint training delivers large explanation gains at a small detection cost. In Table 13(a), joint SFT reduces detection macro-F1 by only 0.68 points relative to Binary-only, while improving explanation precision by 23.08 points and reducing the hallucination rate by 9.44 points. RL adds a further 6.62-point precision gain and 5.26-point hallucination-rate reduction, with only a 0.09-point detection decrease. Binary-only’s highest detection score underscores the need to assess both detection and explanation performance; it does not invalidate the value of joint training.

W4. Self-judge and shared bias.

We select the self-judge based on validated claim-assessment ability, actual training gains, and agreement with independent experts. Its shared origin with the Actor is insufficient to attribute these gains to self-consistency.

(a) RL improves explanation performance beyond the training reward. From SFT to RL, explanation scores on TriDF-B, LOKI-Explain, and FakeClue-Evidence increase from 36.15, 67.69, and 54.84 to 41.01, 73.52, and 61.46, respectively (Tables 10–12). Under identical RL settings, the model trained with the self-judge exceeds the strongest non-self-judge alternative on each metric by 1.56, 2.45, and 2.97 points, respectively (Table 15).

(b) Independent expert evaluation confirms actual explanation gains. From SFT to RL, blinded expert evaluation shows a 7.11-point increase in explanation precision, a 3.48-point decrease in hallucination rate, and a 4.66-point increase in evidence coverage (Table 20(b)). These are not scores assigned by the self-judge to its own outputs and cannot be summarized as the model rewarding its own preferences.

(c) The self-judge’s claim-assessment ability is directly validated. Its judgments of visual support and forensic direction agree with expert consensus at 92.80% and 92.70%, respectively (Table 20(a)). The paper therefore provides independent evidence for both judge quality and final explanation quality. The annotation model’s involvement in review and the self-judge’s shared origin with the Actor do not overturn these results. Attributing the gains to shared bias must also account for the improvements observed by independent experts.

AuthorsOpenReview ↗

Response to Reviewer oFuM (3/3): Explanation Paradigm and Additional Comparisons

W5. Post-hoc explanations and decision faithfulness.

Generating a decision before its explanation follows CLOVER’s task design: binary supervision teaches detection, while evidence supervision teaches verifiable explanations for human review. An explanation need not improve detection or precede the decision to be useful.

(a) Factors such as source coverage in binary training data are an important foundation for improving AI-generated image detection. Table 3(b) validates the effect of source coverage on detection performance. In Table 13(a), Binary-only already achieves 93.12 average macro-F1. By comparison, an “explain-then-decide” approach is constrained by the availability of high-quality explanation annotations, and its detection benefits still require empirical validation.

(b) “Decide-then-explain” combines detection capability with verifiable evidence output. CLOVER uses large-scale binary data for source coverage and limited evidence annotations for explanation learning. Joint SFT reduces detection macro-F1 by only 0.68 points, while increasing explanation precision by 23.08 points and reducing the hallucination rate by 9.44 points (Table 13(a)). Independent expert evaluation further confirms the explanation gains from RL (Table 20(b)).

(c) Verifiable explanations provide evidence for practical review. Reviewers can inspect image evidence and check the model’s judgment instead of receiving only a real/fake label. The practical value of an explanation depends on the quality of its evidence, not its generation order.

W6. Additional baseline comparisons.

The absence of comparisons with closed-source teachers or methods lacking available implementations does not invalidate the existing performance comparisons and controlled experiments. These comparisons are constrained by API evaluation costs and reproducibility resources, respectively.

(a) Outperforming the closed-source teachers is not a premise of our contribution. GPT-5.5 and Claude Opus 4.8 are used offline to construct a limited set of evidence annotations. Evaluating them as detectors across all benchmarks requires a separate API budget. Our goal is to train a locally deployable 8B detection-and-explanation model with limited evidence supervision. Controlled ablations and independent expert evaluation already support the training design and explanation gains; full teacher evaluation is not a prerequisite for these conclusions.

(b) The other three methods currently lack public resources directly usable for a common evaluation. Based on our checks, the official FakeXplainer repository remains empty; PAVE states that code and data will be released after publication; and we could not find a public implementation or model weights for Defake-o3. We will expand the related-work discussion to clarify methodological differences and resource availability. Requesting fuller literature coverage is reasonable, but missing comparisons with methods that lack reproducibility resources is not, by itself, grounds for rejecting the existing experimental conclusions.

Reviewer_gJ92OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a training framework for explainable AI‑generated image detection. The work addresses two critical real‑world pain points: (1) expensive fine‑grained evidence annotations; (2) missing valid visual observations.

The proposed framework suggests to expand detection coverage through binary supervision and advances explanation learning through open-ended evidence verification. Atomic evidence reinforcement learning uses a self-judge to assess claims against the image and rewards distinct supported observations. A frozen self‑judge verifies each atomic claim against input images, and a group‑judge merges semantically overlapping observations to avoid rewarding redundant paraphrases. No ground‑truth reference explanations are required for RL training. Experimental results showed that CLOVER‑SFT achieves good detection performance using an evidence corpus of only ~10% size relative to binary data. Subsequent RL further boosts explanation precision and reduces hallucination rates while largely retaining detection accuracy.

Critical strengths

The proposed training model employs a two‑stage atomic‑evidence RL design without reference rationales. Most prior explainable‑detection RL relies on human‑written reference explanations or region‑level annotations. Instead, this work leverages a self‑judge to perform per‑claim visual verification directly against input images. The semantic grouping mechanism prevents rewarding repeated paraphrases of identical visual cues. The ordered reward schedule (coverage first, then faithfulness) is well‑justified: the model first learns to find candidate observations before being taught to filter unreliable ones.

Critical weaknesses

The model can output abstention notice when trustworthy visual evidence cannot be found. The ablation shows disabling abstention increases hallucination rates. However, the paper does not quantitatively analysis the trade‑off between abstention rate and downstream usability, e.g., recall of usable evidence under different operating points. The main comparison is against prior end‑to‑end explainable detectors. It would be insightful to include additional baselines such as vanilla RL without semantic grouping, or multi‑objective RL that directly optimizes both detection and explanation simultaneously, beyond the existing internal ablations.

Questions for authors

How does the traditional image processing ( heavy compression, screenshot and other real‑world image corruptions ) may affect both detection and evidence reasoning for CLOVER? How abstention rate changes with different operating thresholds, and how much valid forensic evidence is discarded when the model outputs Notice.

2 comments
AuthorsOpenReview ↗

Response to Reviewer gJ92: Abstention, RL Controls, and Corruption Robustness

W1. Abstention and usable evidence.

The paper already quantifies both the benefits and costs of abstention.

(a) Notice withholds an explanation, not the authenticity decision. It is generated directly rather than triggered by a confidence threshold, so threshold sweeping does not match the current inference mechanism. Table 22 instead tests the Notice reward under fixed inference settings.

(b) Compared with the Notice-penalty control, the full model increases Notice rate from 0% to 17.89%, but supported-image coverage decreases only from 70.96% to 70.43%. Precision improves by 3.57 points and hallucination rate drops by 8.66 points: a small coverage cost for substantially more reliable evidence.

(c) Expert assessment finds usable evidence in 75 of 571 Notice images (13.13%), representing 2.35% of all evaluated fake images (Table 23). The other cases have insufficient cues, ambiguous forensic direction, or only out-of-scope candidate explanations. This directly measures how often abstention overlooks usable evidence.

W2. Ungrouped RL and joint optimization.

The paper includes ungrouped RL and joint detection–explanation optimization controls.

(a) Under matched Stage II budgets, semantic grouping outperforms Claim-wise RL by 4.73 precision points, reduces hallucination rate by 4.16 points, and lowers duplicate rate from 23.33% to 4.52%, with nearly unchanged detection macro-F1 (92.33/92.35, Table 16).

(b) Both RL stages optimize detection and explanation; the stages change the evidence reward from coverage to faithfulness. Removing binary rewards reduces detection macro-F1 from 92.35 to 90.77; removing evidence rewards reduces explanation precision from 61.46 to 55.41 (Table 18(a)).

(c) Direct two-task RL (“binary + explainable detection”) achieves detection macro-F1/precision of 92.34/59.47. At fixed binary-task exposure, the full three-task model achieves 92.35/61.46, validating the additional benefit of label-conditioned explanation (Table 18(b)).

W3. Image corruptions.

Additional experiments cover all five detection and three explanation benchmarks, quantifying detection loss, evidence loss, and changes in abstention and hallucinations.

(a) Protocol. We use the original CLOVER-RL checkpoint and fixed inference settings, without adaptation or threshold tuning. Corruptions are JPEG Q70/Q30, 0.5× resizing, Gaussian blur (σ=1.5), and screenshots. Screenshots use a fixed browser, 100% page zoom, DPR=1, and aspect-preserving display at a 512-px long edge; only the image region is captured as PNG.

(b) Detection. Within-benchmark aggregation is unchanged; the average weights all five benchmarks equally.

ConditionGenImageLOKI-DetSoFake-OODFakeBenchAIGIBenchAverage
Original97.9383.0093.7096.1091.0192.35
JPEG Q7096.8081.6092.2094.8089.9091.06
JPEG Q3093.1077.2087.9091.3085.8087.06
Resize 0.5×96.1080.4091.1094.0088.6090.04
Gaussian blur94.5078.3089.2092.4087.1088.30
Screenshot95.7079.6090.5093.6088.1089.50

Macro-F1 (%).

JPEG Q70, resizing, and screenshots reduce average macro-F1 by 1.29, 2.31, and 2.85 points; blur and JPEG Q30 cause larger drops of 4.05 and 5.29 points.

(c) Explanations. TriDF retains the original reference sets to measure original-evidence retention. LOKI-Explain uses corrupted full images and corresponding crops extracted after full-image corruption. FakeClue-Evidence verifies claims against corrupted images. Metrics are reported separately, without cross-benchmark averaging.

ConditionTriDF-A F₀.₅ ↑TriDF-B F₀.₅ ↑LOKI Overall ↑FC Precision ↑FC Hall. Rate ↓FC Evi@Img ↑FC Notice Rate
Original48.6241.0173.5261.469.2970.4317.89
JPEG Q7047.1039.6072.2060.109.8068.8019.30
JPEG Q3041.3034.5066.8055.2011.6063.1026.40
Resize 0.5×45.0037.9070.3058.4010.2066.9022.20
Gaussian blur43.1036.0068.2056.7010.9064.8024.30
Screenshot44.6037.2069.6057.8010.6065.9023.20

FC: FakeClue-Evidence. Notices and empty explanations remain in its evaluation denominator.

JPEG Q30 reduces supported-image coverage by 7.33 points, increases Notice rate by 8.51 points, and increases hallucination rate by 2.31 points. The model abstains more often under severe corruption, but additional evidence loss and unreliable claims remain.

AuthorsOpenReview ↗

Follow-up to Reviewer gJ92

Dear Reviewer gJ92,

As we have not yet heard back from you, we would appreciate your review of our response and additional experiments, and ask that you reconsider your rating in light of this evidence.

If any concerns remain, please let us know. We welcome further discussion and will respond promptly.

Thank you for your time.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Discussion Summary and Final Follow-up

Dear Reviewers and Area Chair,

As the discussion period draws to a close, we thank all reviewers for their thoughtful feedback. The reviews recognized the supervision design for source and scene coverage (oFuM), and claim-level verification and semantic grouping for explanation learning (gJ92, oFuM, 1vSo). We particularly thank 1vSo for following up and acknowledging that our response addressed their concerns.

We summarize below how our responses and additional experiments address the main concerns. Full results and protocols are provided in our General Response and individual replies.

1. Data organization and methodological contribution — oFuM and 1vSo. CLOVER expands detection coverage through binary labels, allocates limited evidence annotations across diverse scenes, and reuses them for label-conditioned explanation.

  • Fixed-budget coverage experiments validate this allocation of supervision.
  • With data, backbone, and total training presentations fixed, the label-conditioned task improves explanation precision by 3.44 points, directly demonstrating a methodological gain before RL.

2. Generalization to unseen generators — oFuM and 1vSo. We added both unseen-family evaluation and retraining with target families removed.

  • On Z-Image, HunyuanImage-3.0, and BAGEL, excluded from all task training and model selection, CLOVER-SFT achieves 80.10% average macro-F1, versus 76.27% / 77.73% for D³ / FakeVLM retrained on matched data.
  • After removing the Nano Banana and GPT Image families from all task-training stages and retraining with matched data volume and budgets, CLOVER-RL achieves 76.10% macro-F1 on Nano-banana and 84.00% fake-image recall on GPT-Image2.

3. Self-judge utility and independent validation — oFuM. The self-judge’s role is to provide effective training signals; complete absence of bias is not a prerequisite.

  • Training utility: Under identical RL settings, the self-judge yields stronger downstream explanation performance than the compared alternative judges. Independent, blinded expert assessment confirms that SFT-to-RL training improves precision by 7.11 points and reduces hallucination rate by 3.48 points.
  • Judgment quality: On the same 2,000 claims, the self-judge achieves 92.40% / 92.00% expert agreement on visual support / forensic direction, the highest among the evaluated judges.

4. Abstention and real-world evaluation — gJ92 and 1vSo. We quantified explanation availability and performance under practical image conditions.

  • Notice retains the authenticity decision while withholding an explanation. Relative to the Notice-penalty control, the full model reduces hallucination rate by 8.66 points, with supported-image coverage nearly unchanged (70.96% to 70.43%).
  • On 1,000 independent real-world images, CLOVER-RL’s false-positive rate is 9.70%, versus 18.20% / 24.60% for Qwen3-VL-8B / FakeVLM. Compression, resizing, blur, and screenshot tests further quantify effects on detection, evidence quality, and abstention.

We hope this summary helps the reviewers and Area Chair weigh the recognized strengths alongside the additional evidence. We would appreciate final feedback from gJ92 and oFuM on our responses and latest results, and welcome any remaining questions before the discussion closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.