{"id":"ccd2dc94-6d0c-4e3f-ad2c-57c028a1f76a","arxiv_id":"2607.14684","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"GlobalForge improves AI-generated image detection under real-world degradation by suppressing local shortcuts and enforcing long-range structural reasoning, outperforming prior state-of-the-art by 5.89% average balanced accuracy on eight in-the-wild benchmark groups.","lead":"This computer-vision paper introduces GlobalForge, an AI-image detector that suppresses local artifacts and forces decisions to rely on long-range structure, reporting about 5.89% higher balanced accuracy than prior state-of-the-art on in-the-wild benchmarks. It also releases a new benchmark, RealDeg-Bench, for testing detection after realistic image degradations like compression and blur.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 5.89% gain over DDA is confounded by backbone: the same-backbone Ours-d2 variant is only +3.19%, and no DINOv3-based DDA or no-module DINOv3 baseline is provided to isolate the module contribution.","rationale":"The reader’s weakest assumption identified exactly the backbone confound in the headline comparison, and my reading agrees. This is the single most load-bearing concern because the paper’s central claim is framed as a SOTA improvement of 5.89% over DDA on in-the-wild benchmarks, and the only same-backbone comparison (Ours-d2 vs. DDA) reduces the gain to 3.19%. The paper does provide internal evidence that the modules matter: Table 3 shows a plain fine-tuned baseline at 80.21 in-the-wild vs. 85.93 for the full model, and the RealDeg-Bench ablation shows similar gains, so I would not reject the method. But the exact headline number is not yet cleanly attributable to the proposal rather than to the newer DINOv3 backbone. A DINOv3-L DDA run is a feasible, concrete control that would settle this. Since the reader already issued a conditional verdict, no adjustment is needed; the condition should be that this control is performed before the 5.89% headline is taken at face value.","tokens_in":18748,"tokens_out":3944,"duration_ms":42817,"concrete_test":"Retrain DDA using DINOv3-L with the same LoRA adaptation, DDA-aligned training data, and evaluation protocol as Ours-d3, and also train a plain DINOv3-L finetune with LIB/GSR/DCS removed (the Table 3 baseline on DINOv3). Evaluate both on the same 17 in-the-wild subsets and compute the 8-group Avg BAcc, ideally with multiple seeds and error bars. If the DINOv3-L DDA or plain baseline reaches within about 1.5 points of 85.93, the headline gain is largely a backbone artifact; if it remains near 80–82, the modules carry the gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper’s headline quantitative claim—Table 2’s Avg BAcc of 85.93% for Ours-d3 vs. 80.04% for DDA, a +5.89% gain—is the load-bearing evidence for “improves … by 5.89% over the previous state-of-the-art.” The problem is that Ours-d3 uses DINOv3-L while DDA is evaluated with officially released weights that, per the paper’s own setup, do not use DINOv3. The backbone-controlled variant Ours-d2 (DINOv2-L, the same family as DDA) reaches 83.23%, only +3.19% over DDA. Thus at least 2.70 points of the headline gain (roughly half) coincides with the backbone change, even ignoring that Ours-d2 and Ours-d3 also differ in epochs, learning rate, and batch settings (Appendix D). Without a DINOv3-L version of DDA, or a plain DINOv3-L no-module baseline evaluated on the same 8 groups, the 5.89% cannot be attributed to LIB/GSR/DCS. The ablations in Table 3 do show that each module helps relative to a plain fine-tuned backbone (baseline 80.21 in-the-wild vs. full 85.93), which supports the mechanism internally, but they do not quantify the backbone contribution to the headline comparison. This is a correctness risk for the claim “5.89% over previous SOTA,” not for the existence of a robustness gain.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that existing AI-generated image detectors fail under real-world degradations because they rely on fragile local artifacts, and proposes GlobalForge, which suppresses local information via a Local Information Bottleneck (LIB), forces long-range evidence gathering via Global Structural Reasoning (GSR), and aligns clean/degraded representations with a degradation-aware contrastive loss (DCS). The authors introduce RealDeg-Bench, a compound-degradation benchmark, and report that GlobalForge improves average BAcc by 5.89% over DDA on 8 in-the-wild benchmark groups and reaches 85.81% average BAcc on RealDeg-Bench. Extensive experiments, ablations, and sensitivity analyses are included.","tokens_in":19231,"tokens_out":9653,"duration_ms":96518,"significance":"If the reported gains hold up, the work is a useful step in a practically important direction: robustness of AI-generated image detectors under realistic propagation. RealDeg-Bench is a sensible controlled benchmark that addresses a real gap, and the explicit release of code and benchmark details supports reproducibility. The paper also makes a conceptually interesting design choice — actively suppressing local shortcuts rather than defending them. The ablations in Table 3 and the alternative analysis in Table 4 provide support for the mechanism. The main weaknesses are attribution of the headline number to the proposed modules versus the backbone and training configuration, and the absence of variance estimates. The paper does not contain formal proofs, but its claims are empirical and falsifiable.","major_comments":[{"comment":"The headline '+5.89% over DDA' uses Ours-d3 (DINOv3-L) against DDA's official released weights, which do not use DINOv3. Table 3 does provide a plain fine-tuned no-module baseline (80.21 in-the-wild) that is essentially tied with DDA (80.04), so I do not think the full gain is purely a backbone effect. However, the gap between Ours-d3 (85.93) and the DINOv2-L variant Ours-d2 (83.23) is 2.70 points, and Appendix D shows those variants also differ in patch size, epochs, learning rate, batch organization, and GPU count. Thus the incremental contribution of LIB/GSR/DCS relative to DDA's pipeline is not cleanly isolated. Please add a DINOv3-L-based DDA run or otherwise control for the backbone, and explicitly separate the 'method + backbone' gain from the 'modules only' gain. At minimum, the 5.89% claim should be qualified as GlobalForge-with-DINOv3-L versus DDA-with-DINOv2-L.","section":"§6.3, Table 2, Appendix D"},{"comment":"All performance numbers are point estimates from a single run, with no error bars, confidence intervals, or multi-seed variation. Several load-bearing comparisons rest on differences of 1–2 BAcc points, e.g., Table 3's w/o-DCS versus Full (83.94 vs 85.93 in-the-wild; 82.61 vs 85.81 on RealDeg-Bench), and some Table 1 rows where Ours-d2 and Ours-d3 are close. Without repeated seeds or a paired statistical test, the component-wise necessity and some benchmark-level gaps are not established. Please add multiple seeds for the main variants and ablations, or report confidence intervals based on subset bootstrapping.","section":"§6.2, Tables 1–3"},{"comment":"The paper's central conceptual claim is that robustness comes from 'global structural coherence' rather than local artifacts. The architecture enforces this by masking local attention and suppressing local features, and the alternative analysis in Table 4 shows that a global-attention variant performs worse. However, there is no direct measurement that the learned decisions actually rely on long-range structure — for example, a perturbation study that compares sensitivity to distant versus adjacent regions, or a statistic over token-to-token attention distances. As written, 'global structural cue' is an interpretation of why the architecture succeeds, not an independently verified property. I recommend either adding such an experiment or softening the 'root cause' phrasing in the abstract and introduction.","section":"§3 and §4.2"}],"minor_comments":[{"comment":"The 5.89% headline should state explicitly that it compares the DINOv3-L variant against DDA's official weights; otherwise readers may infer a controlled module-only comparison.","section":"Abstract, §1"},{"comment":"The table lists 17 subset scores across 8 groups, but the group structure is not fully legible in the printed header. A subtable or explicit group-to-subset mapping would improve reproducibility.","section":"Table 2"},{"comment":"The notation 'RealDeg-1Step' is used in the caption but formally defined only in Appendix D. Define it in the caption or use 'RealDeg-Bench compound N=1' for clarity.","section":"Table 4, Appendix D"},{"comment":"The denominator sums over j without specifying whether the index range includes the positive pair in both terms. State explicitly that j runs over the full batch including the positive; otherwise the loss expression is ambiguous.","section":"Eq. (6)"},{"comment":"The training differences between ours-d2 and ours-d3 (epochs 10/8 vs 8/6, LR 4e-4 vs 5e-5, batch/GPU 128/16 vs 128/4, patch sizes 14 vs 16) should be highlighted in the main text, since readers will otherwise interpret the Ours-d2/Ours-d3 gap as purely a backbone effect.","section":"Appendix D, Table 7"},{"comment":"The x-axis for JPEG quality factor is drawn with high quality on the left and low quality on the right, which is opposite to the horizontal direction used in most other panels. A consistent monotonic direction would avoid misreading.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The core empirical contribution appears credible, and the benchmark is a useful addition. The main risk is over-claiming the attribution of the head-line number to the proposed modules; this is fixable with a DINOv3-L control and variance reporting. I do not see integrity or novelty-disclosure concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: the paper is worth refereeing. The core claim—that AIGI detectors collapse because they latch onto local artifacts, and that you can force the model onto global structural cues—is supported by a controlled diagnosis and a clean module-level ablation. What's new: LIB (feature-domain local suppression), GSR (masked attention that blocks the local window), and a degradation-aware contrastive loss, plus RealDeg-Bench, a compound-degradation evaluation set. The benchmark alone is a useful contribution: 7 operators, 5-step chains, reproducible seeds and strengths.\n\nThe experiments are extensive and the trend is credible. On RealDeg-Bench and the 8 in-the-wild groups, GlobalForge is consistently ahead, and the strength sweeps show a much flatter decay than DDA or B-Free. The alternative-analysis table (input blur, FFT, global attention) makes a fair comparison and shows the feature-domain design isn't just low-pass filtering.\n\nSoft spots:\n\n1. The headline '5.89% over previous SOTA' is a system-level comparison: Ours-d3 uses DINOv3-L while DDA is evaluated with its official weights. That alone could be a real confound. But the ablation in Table 3 gives a plain fine-tuned DINOv3-L baseline at 80.21 BAcc, nearly identical to DDA's 80.04, and the full model is 85.93. So the modules account for roughly 5.7 points on the same backbone; the backbone contribution looks small. The stress-test worry that half the gain is backbone is not supported by the paper's own numbers. Still, the authors should report that no-module baseline prominently in the main comparison, and ideally evaluate a DINOv3-based DDA, because as written the reader has to dig into Table 3 to settle the confound.\n\n2. Single run, seed 0, no error bars. For a robustness paper, one seed is thin. The improvement is large enough that I'd expect it to survive re-seeding, but a reviewer should ask for at least 3 seeds.\n\n3. Ours-d2 vs Ours-d3 differ in epochs, LR, batch size, so the 'DINOv3' variant is really also a different training schedule. The internal ablation controls for that only for d3; the d2 vs d3 comparison should not be read as a clean backbone study.\n\n4. Baseline protocol mixes official and retrained weights. Appendix C gives a reasonable rationale and a comparison table, so this is acceptable, but it adds noise to the SOTA claim.\n\nWho this is for: people working on AIGI detection, robustness, or benchmark construction. The mechanism story is not proven by the Grad-CAM figures alone, but the quantitative evidence is there.\n\nRecommendation: send to peer review. Ask for error bars and a same-backbone/same-schedule comparison with DDA, or at least a clear statement that the headline is a system-level result. The paper deserves referee time.","headline":"GlobalForge is a credible robustness-first detector with a new benchmark; the headline SOTA gap is a system-level comparison, but the internal same-backbone ablation mostly addresses the backbone confound, so the fix is reporting, not a redesign.","tokens_in":19688,"tokens_out":3720,"would_cite":true,"duration_ms":36752,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GlobalForge claims that AI-generated-image detection becomes robust to real-world degradation when detectors are steered from fragile local artifacts to degradation-stable global structural coherence, and reports a 5.89% average balanced-ac","keywords":["AI-generated image detection","deepfake detection","degradation robustness","shortcut learning","global structural reasoning","Local Information Bottleneck","contrastive learning","benchmark"],"falsifier":"Train the strongest prior detector with the same newer backbone, fine-tuning recipe, training data, and schedule as GlobalForge and evaluate on the same eight in-the-wild groups; if it reaches or exceeds 85.93% average balanced accuracy, the proposed modules add little beyond a backbone upgrade.","tokens_in":18684,"feed_emoji":"🖼️","tokens_out":6086,"duration_ms":56953,"temperature":0.7,"pith_summary":"This paper argues that AI-generated-image detectors fail under real-world degradation because they rely on local generator artifacts that JPEG, blur, and resize destroy. To fix that, GlobalForge trains detectors to ignore local details and reason over long-range structural coherence: a Local Information Bottleneck suppresses high-frequency components, a Global Structural Reasoning module masks nearby tokens so each token must gather evidence from distant regions, and a contrastive loss keeps features stable under degradation. The paper backs the claim with a new benchmark, RealDeg-Bench, that chains one to five degradation operators, and reports that GlobalForge improves average balanced accuracy on 8 in-the-wild benchmark groups by 5.89% over the previous state of the art, reaching 85.81% on RealDeg-Bench. A sympathetic reader would care because it proposes a mechanism, not just data augmentation, for why detectors break and how to build them to survive social-media transmission.","feed_headline":"Switching AI-image detectors from local to global cues gains 5.89%","feed_subtitle":"GlobalForge plus RealDeg-Bench keeps AI-image detection accurate through JPEG, blur, and multi-step resize chains.","key_machinery":"The framework's load-bearing mechanism is a pair of modules inserted after a frozen pretrained vision transformer. LIB (Local Information Bottleneck) forms a convex combination of deep features with a Gaussian-blurred copy, suppressing high-frequency local components in feature space rather than in pixel space. GSR (Global Structural Reasoning) masks self-attention within a Chebyshev distance of 3, so each token can only aggregate evidence from distant regions; with masking probability 1.0 during training, the model cannot fall back on adjacent-token correlations. A degradation-aware contrastive structural loss (L_DCS) pulls clean and compound-degraded views together in feature space, stabil","core_discovery":"The paper's central claim is that AI-generated-image detectors fail under real-world degradation because they overfit to generator-specific local artifacts, and that actively suppressing local information while forcing long-range structural reasoning produces a cue that survives JPEG, blur, resize, and multi-step chains. In the authors' experiments, this design—trained with a degradation-aware contrastive loss—raises average balanced accuracy on eight in-the-wild benchmark groups to 85.93%, which they report as 5.89 points above the strongest prior detector, and reaches 85.81% on their new RealDeg-Bench. Ablations show each module (LIB, GSR, and the contrastive loss) contributes, and that in","pith_inferences":["Editorial inference: the same local-suppression/global-reasoning recipe could be transferred to other forensic tasks, such as video deepfake detection, where compression chains tend to be deeper; the paper's appendix lists video extension as future work.","Editorial inference: if global structural coherence is the real cue, targeted attacks that distort long-range statistics while leaving local patches intact should be more effective against GlobalForge than local-patch attacks; the paper does not test this.","Editorial inference: the benchmark's with-replacement chain sampling makes repeated re-compression a first-class citizen; a natural community norm would be to report accuracy as a function of chain depth, not only endpoint accuracy."],"forward_implications":["If the claim holds, detectors can be made robust without restoration or augmentation-only defenses; the cue itself is stable under degradation.","Accuracy on clean images need not be sacrificed: the authors report the best average degraded accuracy across six standard benchmarks and near-best clean accuracy.","The failure mode should become graceful rather than cliff-like: performance decays smoothly from one to five chained degradations.","The mechanism should also transfer across generators, because global structure is less tied to a specific generator's local signature.","RealDeg-Bench supplies a fixed 13-condition protocol (clean, seven single degradations, five compound chains) so future detectors can be compared on the same degradation axis."],"fun_headline_variants":["AI detectors gain 5.89% by ditching local cues for global","GlobalForge: 5.89% accuracy gain under real-world degradation","For global robustness, AI-image detector suppresses local noise","Why AI-image detectors fail? Local glitches. Fix? Go global","Local artifacts fool detectors; global structure survives"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that the 5.89% gain comes from the proposed modules, rather than from the stronger newer backbone used in the best GlobalForge variant, is not fully isolated by an explicit backbone-controlled baseline; the only controlled comparison with the older backbone shows a smaller 3.19% gain over the same prior detector.","fun_headline_variants_meta":{"raw":{"variants":["AI detectors gain 5.89% by ditching local cues for global","GlobalForge: 5.89% accuracy gain under real-world degradation","For global robustness, AI-image detector suppresses local noise","Why AI-image detectors fail? Local glitches. Fix? Go global","Local artifacts fool detectors; global structure survives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001301,"raw_usage":{"total_tokens":5159,"prompt_tokens":774,"completion_tokens":4385,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":4297}},"tokens_in":518,"tokens_out":4385,"duration_ms":34403,"temperature":1.0,"reasoning_tokens":4297,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T01:20:53.726597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the strongest prior detector with the same newer backbone, fine-tuning recipe, training data, and schedule as GlobalForge and evaluate on the same eight in-the-wild groups; if it reaches or exceeds 85.93% average balanced accuracy, the proposed modules add little beyond a backbone upgrade.","supporting_citations":[],"review_version":1}