Pith. sign in

REVIEW 4 major objections 1 cited by

An LLM agent guided by expert reliability and image-context profiles can fuse conflicting AI-image detectors and issue auditable authenticity reports.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 19:50 UTC pith:U6C3YPU3

load-bearing objection Solid systems paper on profile-guided LLM fusion for AIGI with real explainability and modular integration; X-Fuse is a valid stress test but partly manufactures the headline margin, so lean on the in-the-wild numbers. the 4 major comments →

arxiv 2603.23115 v2 pith:U6C3YPU3 submitted 2026-03-24 cs.CV

AgentFoX: LLM Agent-Guided Fusion with eXplainability for AI-Generated Image Detection

Yangxin Yu , Yue Zhou , Bin Li , Kaiqing Lin , Haodong Li , Jiangqun Ni , Bo Cao This is my paper
classification cs.CV
keywords AI-generated image detectionmulti-expert fusionLLM agentsexplainable forensicsprobability calibrationclustering profilesconflict resolutionAIGI
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Specialized detectors for AI-generated images often disagree because each is tuned to different artifacts and fails on different images. This paper claims that an LLM agent can resolve those disagreements by consulting two offline knowledge bases: Expert Profiles that calibrate each detector’s trustworthiness, and Clustering Profiles that rank how reliable each detector is on images that share similar semantic, texture, or camera-artifact features. Guided by a fixed forensic guideline, the agent first runs a semantic check, then gathers calibrated signal-level scores, arbitrates conflicts with the profiles, and writes a step-by-step report that states the final verdict and its evidence. On a deliberately high-conflict benchmark assembled from seven public test sets and on real-world in-the-wild collections, the approach outperforms individual detectors and conventional ensembles while remaining modular: a new detector can be added after a single profiling pass without retraining the fusion stage. The result is a detection pipeline that is both more accurate under disagreement and transparent enough for forensic use.

Core claim

AgentFoX establishes that an LLM agent equipped with calibrated Expert Profiles for model-centric reliability and Clustering Profiles for data-centric local reliability can perform sample-specific fusion of heterogeneous AIGI detectors through multi-stage, guideline-driven reasoning, yielding higher accuracy than single experts or fixed ensembles on high-conflict and in-the-wild benchmarks while producing a fully auditable forensic report.

What carries the argument

Expert Profiles (calibrated reliability summaries and executable score mappings for each detector) and Clustering Profiles (per-modality cluster rankings of local detector reliability plus quality scores), which the agent queries inside a structured multi-turn loop to reweight evidence and adjudicate conflicts under a fixed forensic guideline.

Load-bearing premise

The offline Expert and Clustering Profiles built from a limited reference split remain trustworthy guides for reweighting detectors on images from generators, domains, and post-processing never seen during profiling.

What would settle it

Construct a fresh high-conflict test set drawn only from generators and post-processing pipelines absent from the profile-construction data; if AgentFoX accuracy falls to or below the best single expert or simple majority vote while the profiles still assign high local reliability to the failing detectors, the transfer claim is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • New AIGI detectors can be registered by a one-time offline profiling step and used immediately, without retraining any fusion model.
  • Every authenticity decision is accompanied by a human-readable report that cites the evidence, expert agreements, and the exact rationale used to resolve conflicts.
  • Context-aware reweighting lets the system down-weight detectors that are globally strong yet locally unreliable for a given image cluster.
  • The same modular agent-plus-profile pattern can absorb future signal-level tools as they appear, keeping the forensic stack current without architectural rewrites.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The profile-plus-agent pattern could transfer to other multi-sensor forensic settings—audio deepfakes, document forgery, or video reenactment—where specialized models likewise produce conflicting local signals.
  • Heavy JPEG compression or other post-processing that erases the high-frequency cues many experts rely on will force the agent to lean more on semantic analysis; if that analysis is itself weak, overall reliability will drop even with correct profiles.
  • The multi-stage, multi-thousand-token workflow currently favors high-stakes investigation over real-time screening; shorter adaptive expert-selection policies would be a natural next engineering target.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 0 minor

Summary. AgentFoX is an LLM-agent framework for multi-expert AI-generated image (AIGI) detection. Offline, it builds Expert Profiles (calibrated reliability, textual descriptions) and Clustering Profiles (per-modality local reliability rankings over CLIP/SRM/CFA clusters). At inference, a guideline-driven ReAct-style agent runs semantic VLLM analysis, queries calibrated signal-level experts (DRCT, RINE, SPAI, Patch Shuffle), arbitrates conflicts with profile context, and emits an auditable forensic report. The authors introduce X-Fuse, a stratified high-conflict benchmark over 2^{j+1} expert-error/GT vectors, and report strong overall results for AgentFoX-32B on X-Fuse (F1/ACC 0.8087/0.7949) and on in-the-wild WIRA/WildFake, with ablations on expert count and profile modules, plus supplementary seed, JPEG, and cost analyses.

Significance. If the results hold under fairer framing, the work is a useful systems contribution: it treats multi-detector fusion as sample-specific, profile-guided adjudication rather than fixed voting or retrained meta-classifiers, and it couples that with human-readable provenance. Strengths include an explicit high-conflict evaluation design, ablations that move in the expected direction (Tables 3–4), seed-stability and JPEG checks in the supplement, a practical quick-integration path without fusion-stage retraining, and released code. The in-the-wild gains (Table 2) are more persuasive than the X-Fuse headline alone. Explainable multi-expert forensics is a timely direction even if the explanation quality is not yet measured.

major comments (4)
  1. Contribution 3 and §4.2 claim superior accuracy and that AgentFoX “consistently” outperforms individual detectors and conventional ensembles. Table 1 contradicts a uniform reading: on GenImage, ResNet Fusion (0.9478/0.9465) and MLP Fusion (0.9353/0.9359) substantially beat AgentFoX-32B (0.7854/0.7804); similar patterns appear on AIGCDetect. The overall win is driven by harder/high-conflict cells and the stratified mix. Please restate claims as overall/high-conflict and in-the-wild gains, not dataset-uniform superiority, and discuss when learnable fusion remains preferable.
  2. §4.1, Eqs. (9)–(10): X-Fuse is built by stratified sampling over all expert-error/GT conflict vectors so that single-expert accuracy is driven toward ~0.5 and disagreement is maximized. That is a legitimate stress test for conflict resolution, but it also inflates the headline margin of any method that can reweight the same four experts. The paper’s strongest quantitative claim is still anchored on this adversarial distribution. Please (i) explicitly frame X-Fuse as a conflict-resolution stress test rather than a general detection leaderboard, (ii) report non-stratified or natural-mix results for the same sources, and (iii) weight Table 2 (WIRA/WildFake) more heavily in the abstract and conclusion.
  3. A core claimed contribution is explainability via auditable forensic reports (Abstract; contribution 2; §3.2 Stage 4; Fig. 3). The manuscript provides qualitative traces and a guideline prompt (supplement) but no quantitative evaluation of report faithfulness, completeness, or human usefulness (e.g., agreement with expert rationales, ablation of report components, or a small user study). Without that, the “trustworthiness/interpretability” claim is unsupported relative to the detection numbers. Either add a minimal evaluation or narrow the claim to “structured, evidence-linked reports” without asserting enhanced trust.
  4. §3.1–3.2 and the weakest transfer assumption: Expert/Clustering Profiles are fit on a GenImage-derived calibration/clustering split and applied zero-shot elsewhere. Table 4 shows both modules help on X-Fuse, but there is no analysis of cluster-assignment stability or local-ranking correlation on held-out generators/domains (e.g., agreement of R_c rankings between GenImage reference clusters and X-Fuse/WIRA subsets). If local reliability priors mis-transfer, conflict arbitration is under-grounded. Please add a transfer diagnostic or an out-of-reference ablation that freezes vs. re-estimates cluster reliability.

Circularity Check

0 steps flagged

Empirical systems paper with no derivation circularity; profiles are fit on held-out GenImage data and applied zero-shot to non-overlapping tests.

full rationale

AgentFoX is an engineering/systems paper, not a first-principles derivation. Expert Profiles (calibration via ECE/Brier on D_val, Eq. 1–2) and Clustering Profiles (K-means on CLIP/SRM/CFA features with local F1/ACC rankings, Eq. 3–4) are constructed offline on a GenImage training/validation split (Sect. 4.1: 4,800 images, 80/20 for calibration/clustering vs. profile generation). These priors are then used zero-shot by a guideline-driven ReAct agent (Sect. 3.2, Stages 1–4) on X-Fuse and in-the-wild sets that the authors state are hash-deduplicated with no overlap to training/profile data. Ablations (Tables 3–4) show performance rises with more experts and with both profile types, but remains above many SOTA baselines even without profiles (F1 0.7032), so the headline gains are not forced by construction. X-Fuse’s stratified conflict-vector sampling (Eqs. 9–10) is an adversarial evaluation design that can inflate fusion margins relative to natural distributions; that is a validity/selection concern, not circularity of a claimed derivation. No self-definitional equations, fitted-input-as-prediction, load-bearing uniqueness theorems, or ansatz-via-self-citation appear. Score 1 reflects only the ordinary use of authors’ own prior detectors as modular experts, which is not circular.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 4 invented entities

The central claim rests on empirical ML assumptions and engineered constructs rather than free physical constants. Load-bearing pieces are: (1) offline profiles estimated on GenImage-derived splits transfer as reliability priors; (2) multi-modality clustering captures context-dependent expert competence; (3) an LLM constrained by forensic guidelines can adjudicate conflicts without inventing evidence; (4) several design knobs (cluster counts, calibrator choice, expert panel, X-Fuse sampling) chosen by the authors. Invented entities are the profile objects and the agent framework itself; they have no independent existence outside this system.

free parameters (5)
  • Cluster counts K_t (CLIP/SRM/CFA) = 12 / 8 / 10
    Chosen via inertia elbow / peak second derivative on the profile reference split (reported K=12, 8, 10). Directly shapes Clustering Profiles used in conflict resolution.
  • Per-expert calibration mapping F̂_j = argmin ECE among Temperature/Platt/Isotonic/Histogram/Beta
    Selected among five post-hoc calibrators by lowest ECE on D_val; transforms raw scores into the signals the agent trusts.
  • X-Fuse sampling quota n per conflict vector = 15
    Hand-set n=15 (with d=7, j=4) to build the high-conflict stress test that drives the main claim.
  • Profile data split (GenImage train cells) = 4800 images; 80/20 split
    4800 images, 80% calibration/clustering/aggregation, 20% profile generation/model selection; defines all offline priors.
  • LLM decoding controls = temperature=0, seed=42
    Temperature 0 and seed 42 fixed for reported runs; stochasticity still affects long traces (seed study in supplement).
axioms (5)
  • domain assumption Heterogeneous signal-level AIGI detectors have complementary, context-dependent expertise that can be productively fused rather than merely averaged.
    Motivated by Fig. 1 consensus analysis and used throughout the fusion design (Introduction, Sect. 3).
  • domain assumption Clustering images in CLIP, SRM, and CFA feature spaces yields clusters whose local F1/ACC rankings are meaningful reliability priors for new images assigned to those clusters.
    Core of Clustering Profiles (Sect. 3.1); if clusters do not align with failure modes, arbitration is misinformed.
  • domain assumption An LLM agent following a fixed forensic guideline (ReAct-style tool use) can synthesize multi-source evidence and resolve conflicts without unconstrained hallucination.
    Stated as design principle in Sect. 3.2; limitations section later admits residual reasoning drift.
  • domain assumption Post-hoc probability calibration (ECE-selected) makes heterogeneous detector scores comparable enough for agent-guided weighting.
    Sect. 3.1 Expert Profiles; standard calibration practice treated as sufficient for forensic fusion.
  • domain assumption X-Fuse samples drawn from official test splits of seven public datasets with hash deduplication do not overlap profile/training data and fairly stress fusion methods.
    Sect. 4.1 construction claim; evaluation validity depends on it.
invented entities (4)
  • Expert Profiles (T_desc, T_quality, calibrated tool F̂) no independent evidence
    purpose: Provide model-centric reliability text and calibrated scores for agent fusion.
    Structured object defined in Eq. (2); exists only as part of this framework.
  • Clustering Profiles (per-modality reliability rankings and quality summaries) no independent evidence
    purpose: Supply data-centric, cluster-local expert rankings for conflict resolution.
    Defined in Eq. (4); no external validation outside the paper’s pipeline.
  • AgentFoX multi-stage forensic agent + guidelines G no independent evidence
    purpose: Orchestrate semantic analysis, expert querying, arbitration, and report generation.
    The proposed system; performance is the paper’s own evaluation of this construct.
  • X-Fuse high-conflict benchmark no independent evidence
    purpose: Stress-test multi-expert fusion under stratified disagreement patterns.
    Author-curated sampling protocol (Sect. 4.1); not an independent community standard.

pith-pipeline@v1.1.0-grok45 · 23879 in / 4095 out tokens · 51307 ms · 2026-07-13T19:50:21.645785+00:00 · methodology

0 comments
read the original abstract

The realism of AI-generated images (AIGI) poses increasing challenges for reliable forensic detection, where heterogeneous expert detectors may produce conflicting predictions across diverse generative sources and post-processing conditions. Existing multi-expert fusion methods rely on fixed rules or learned fusion strategies, offering limited ability to assess sample-specific reliability, execute rigorous adjudication of conflicts, and provide evidence-grounded explanations. We propose AgentFoX, an LLM-driven agentic multi-expert framework for AIGI detection that employs a command-and-reasoning core to perform evidence fusion. Following predefined guidelines, the core coordinates designated subtasks to collect semantic and signal-level evidence, reason over structured contexts to determine authenticity, and generate an auditable report for explainability. During this process, Expert Profiles are constructed for model-centric reliability assessment, while Clustering Profiles are built for data-centric contextual analysis, jointly establishing evidence contexts for conflict resolution. Extensive evaluations across diverse benchmarks demonstrate the robustness and generalizability of AgentFoX under complex conditions.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Veritas++: Value-aware On-Policy Distillation for Perception-Enhanced AIGI Detection

    cs.CV 2026-07 conditional novelty 6.0

    Strengthening fine-grained, semantic-anomaly, and pixel-level perception with verifiable rewards, then value-aware on-policy self-distillation, improves generalizable MLLM AI-image detection and adaptation.

Reference graph

Works this paper leans on

49 extracted references · 10 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Finding compact and well-separated clusters: Clustering us- ing silhouette coefficients.Pattern Recognition, 135:109144,

    Adil M Bagirov, Ramiz M Aliguliyev, and Nargiz Sultanova. Finding compact and well-separated clusters: Clustering us- ing silhouette coefficients.Pattern Recognition, 135:109144,

  2. [2]

    Drct: diffusion reconstruction contrastive training towards universal detection of diffusion generated images

    Baoying Chen, Jishen Zeng, Jianquan Yang, and Rui Yang. Drct: diffusion reconstruction contrastive training towards universal detection of diffusion generated images. InPro- ceedings of the 41st International Conference on Machine Learning. JMLR.org, 2024. 1, 7, 8

    2024

  3. [3]

    Dual data alignment makes ai- generated image detector easier generalizable.arXiv preprint arXiv:2505.14359, 2025

    Ruoxin Chen, Junwei Xi, Zhiyuan Yan, Ke-Yue Zhang, Shuang Wu, Jingyi Xie, Xu Chen, Lei Xu, Isabel Guan, Taiping Yao, et al. Dual data alignment makes ai- generated image detector easier generalizable.arXiv preprint arXiv:2505.14359, 2025. 2

    arXiv 2025

  4. [4]

    X2-dfd: A framework for ex- plainable and extendable deepfake detection.arXiv preprint arXiv:2410.06126, 2024

    Yize Chen, Zhiyuan Yan, Guangliang Cheng, Kangran Zhao, Siwei Lyu, and Baoyuan Wu. X2-dfd: A framework for ex- plainable and extendable deepfake detection.arXiv preprint arXiv:2410.06126, 2024. 2

    Pith/arXiv arXiv 2024

  5. [5]

    Davies and Donald W

    David L. Davies and Donald W. Bouldin. A cluster separa- tion measure.IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-1(2):224–227, 1979. 4

    1979

  6. [6]

    Rich models for steganal- ysis of digital images.IEEE Transactions on information Forensics and Security, 7(3):868–882, 2012

    Jessica Fridrich and Jan Kodovsky. Rich models for steganal- ysis of digital images.IEEE Transactions on information Forensics and Security, 7(3):868–882, 2012. 4

    2012

  7. [7]

    Verification of forecasts expressed in terms of probability.Monthly weather review, 78(1):1–3,

    W Brier Glenn et al. Verification of forecasts expressed in terms of probability.Monthly weather review, 78(1):1–3,

  8. [8]

    A bias-free training paradigm for more general ai-generated image de- tection

    Fabrizio Guillaro, Giada Zingarini, Ben Usman, Avneesh Sud, Davide Cozzolino, and Luisa Verdoliva. A bias-free training paradigm for more general ai-generated image de- tection. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 18685–18694, 2025. 2

    2025

  9. [9]

    On calibration of modern neural networks

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. InInternational conference on machine learning, pages 1321–1330. PMLR,

  10. [10]

    Wildfake: A large-scale chal- lenging dataset for ai-generated images detection.arXiv preprint arXiv:2402.11843, 2024

    Yan Hong and Jianfu Zhang. Wildfake: A large-scale chal- lenging dataset for ai-generated images detection.arXiv preprint arXiv:2402.11843, 2024. 7, 8

    Pith/arXiv arXiv 2024

  11. [11]

    UniShield: An adaptive multi-agent framework for unified forgery image detection and localization, 2025

    Qing Huang, Zhipei Xu, Xuanyu Zhang, and Jian Zhang. UniShield: An adaptive multi-agent framework for unified forgery image detection and localization, 2025. 3

    2025

  12. [12]

    Interpretable and reliable detection of ai-generated images via grounded rea- soning in mllms.arXiv preprint arXiv:2506.07045, 2025

    Yikun Ji, Hong Yan, Jun Lan, Huijia Zhu, Weiqiang Wang, Qi Fan, Liqing Zhang, and Jianfu Zhang. Interpretable and reliable detection of ai-generated images via grounded rea- soning in mllms.arXiv preprint arXiv:2506.07045, 2025. 2

    Pith/arXiv arXiv 2025

  13. [13]

    Any-resolution ai- generated image detection by spectral learning

    Dimitrios Karageorgiou, Symeon Papadopoulos, Ioannis Kompatsiaris, and Efstratios Gavves. Any-resolution ai- generated image detection by spectral learning. InProceed- ings of the Computer Vision and Pattern Recognition Con- ference, pages 18706–18717, 2025. 1, 7, 8

    2025

  14. [14]

    Deeclip: A robust and generalizable transformer-based framework for detecting ai-generated images.arXiv preprint arXiv:2504.19876, 2025

    Mamadou Keita, Wassim Hamidouche, Hessen Bougueffa Eutamene, Abdelmalik Taleb-Ahmed, and Abdenour Ha- did. Deeclip: A robust and generalizable transformer-based framework for detecting ai-generated images.arXiv preprint arXiv:2504.19876, 2025. 7, 8

    Pith/arXiv arXiv 2025

  15. [15]

    Leveraging rep- resentations from intermediate encoder-blocks for synthetic image detection

    Christos Koutlis and Symeon Papadopoulos. Leveraging rep- resentations from intermediate encoder-blocks for synthetic image detection. InEuropean Conference on Computer Vi- sion, pages 394–411. Springer, 2024. 1, 2, 7, 8

    2024

  16. [16]

    Beta cal- ibration: a well-founded and easily implemented improve- ment on logistic calibration for binary classifiers

    Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta cal- ibration: a well-founded and easily implemented improve- ment on logistic calibration for binary classifiers. InArtifi- cial intelligence and statistics, pages 623–631. PMLR, 2017. 4

    2017

  17. [17]

    Agent4FaceForgery: Multi-agent LLM framework for realistic face forgery detection, 2025

    Yingxin Lai, Zitong Yu, Jun Wang, Linlin Shen, Yong Xu, and Xiaochun Cao. Agent4FaceForgery: Multi-agent LLM framework for realistic face forgery detection, 2025. 3

    2025

  18. [18]

    Bridging the gap between ideal and real-world evaluation: Benchmarking ai-generated image detection in challenging scenarios

    Chunxiao Li, Xiaoxiao Wang, Meiling Li, Boming Miao, Peng Sun, Yunjian Zhang, Xiangyang Ji, and Yao Zhu. Bridging the gap between ideal and real-world evaluation: Benchmarking ai-generated image detection in challenging scenarios. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 20379–20389, 2025. 3 9

    2025

  19. [19]

    Is artificial intelligence gen- erated image detection a solved problem?arXiv preprint arXiv:2505.12335, 2025

    Ziqiang Li, Jiazhen Yan, Ziwen He, Kai Zeng, Weiwei Jiang, Lizhi Xiong, and Zhangjie Fu. Is artificial intelligence gen- erated image detection a solved problem?arXiv preprint arXiv:2505.12335, 2025. 7, 8

    arXiv 2025

  20. [20]

    Standing on the shoulders of giants: Reprogram- ming visual-language model for general deepfake detection

    Kaiqing Lin, Yuzhen Lin, Weixiang Li, Taiping Yao, and Bin Li. Standing on the shoulders of giants: Reprogram- ming visual-language model for general deepfake detection. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 5262–5270, 2025. 2

    2025

  21. [21]

    Seeing before reasoning: A unified frame- work for generalizable and explainable fake image detection

    Kaiqing Lin, Zhiyuan Yan, Ruoxin Chen, Junyan Ye, Ke- Yue Zhang, Yue Zhou, Peng Jin, Bin Li, Taiping Yao, and Shouhong Ding. Seeing before reasoning: A unified frame- work for generalizable and explainable fake image detection. arXiv preprint arXiv:2509.25502, 2025. 2

    arXiv 2025

  22. [22]

    Forgery-aware adaptive transformer for generalizable synthetic image detection

    Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Jingdong Wang, and Yao Zhao. Forgery-aware adaptive transformer for generalizable synthetic image detection. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 10770–10780, 2024. 2, 12, 13

    2024

  23. [23]

    Color filter arrays: Design and performance analysis.IEEE Transactions on Consumer electronics, 51(4):1260–1267, 2005

    Rastislav Lukac and Konstantinos N Plataniotis. Color filter arrays: Design and performance analysis.IEEE Transactions on Consumer electronics, 51(4):1260–1267, 2005. 4

    2005

  24. [24]

    What is real anymore? an ai/ml image dataset using authenticity validation and traceable origins for every data instance

    Andrew McDonald. What is real anymore? an ai/ml image dataset using authenticity validation and traceable origins for every data instance. InWorkshop on Datasets and Evaluators of AI Safety, 2025. 7, 8

    2025

  25. [25]

    Predicting good probabilities with supervised learning

    Alexandru Niculescu-Mizil and Rich Caruana. Predicting good probabilities with supervised learning. InProceedings of the 22nd International Conference on Machine Learning, page 625–632, New York, NY , USA, 2005. Association for Computing Machinery. 4

    2005

  26. [26]

    Measuring calibration in deep learning

    Jeremy Nixon, Michael W Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. Measuring calibration in deep learning. InCVPR workshops, 2019. 4

    2019

  27. [27]

    Towards uni- versal fake image detectors that generalize across genera- tive models

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards uni- versal fake image detectors that generalize across genera- tive models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480– 24489, 2023. 12, 13

    2023

  28. [28]

    Gpt-4o system card, 2024

    OpenAI. Gpt-4o system card, 2024. 6, 7, 8

    2024

  29. [29]

    Community forensics: Using thousands of generators to train fake image detectors

    Jeongsoo Park and Andrew Owens. Community forensics: Using thousands of generators to train fake image detectors. InProceedings of the Computer Vision and Pattern Recogni- tion Conference, pages 8245–8257, 2025. 7, 8

    2025

  30. [30]

    Probabilistic outputs for support vector ma- chines and comparisons to regularized likelihood methods

    John Platt et al. Probabilistic outputs for support vector ma- chines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61–74, 1999. 4

    1999

  31. [31]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 4

    2021

  32. [32]

    Frequency-aware deepfake de- tection: improving generalizability through frequency space domain learning

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake de- tection: improving generalizability through frequency space domain learning. InProceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Con- ference on Innovative Applications of Artificial Intelligence and Fourt...

    2024

  33. [33]

    Rethinking the up-sampling op- erations in cnn-based generative network for generalizable deepfake detection

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling op- erations in cnn-based generative network for generalizable deepfake detection. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 28130–28139, 2024. 12, 13

    2024

  34. [34]

    C2p-clip: Inject- ing category common prompt in clip to enhance generaliza- tion in deepfake detection

    Chuangchuang Tan, Renshuai Tao, Huan Liu, Guanghua Gu, Baoyuan Wu, Yao Zhao, and Yunchao Wei. C2p-clip: Inject- ing category common prompt in clip to enhance generaliza- tion in deepfake detection. InProceedings of the AAAI Con- ference on Artificial Intelligence, pages 7184–7192, 2025. 2, 7, 8

    2025

  35. [35]

    Cnn-generated images are surprisingly easy to spot

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695–8704, 2020. 2

    2020

  36. [36]

    Spot the fake: Large multimodal model-based synthetic image detection with artifact explanation.arXiv preprint arXiv:2503.14905, 2025

    Siwei Wen, Junyan Ye, Peilin Feng, Hengrui Kang, Zichen Wen, Yize Chen, Jiang Wu, Wenjun Wu, Conghui He, and Weijia Li. Spot the fake: Large multimodal model-based synthetic image detection with artifact explanation.arXiv preprint arXiv:2503.14905, 2025. 2, 5, 7, 8

    arXiv 2025

  37. [37]

    Avatarshield: Visual reinforcement learning for human-centric video forgery detection.arXiv preprint arXiv:2505.15173, 2025

    Zhipei Xu, Xuanyu Zhang, Xing Zhou, and Jian Zhang. Avatarshield: Visual reinforcement learning for human-centric video forgery detection.arXiv preprint arXiv:2505.15173, 2025. 2, 5

    arXiv 2025

  38. [38]

    A sanity check for ai-generated image detection.arXiv preprint arXiv:2406.19435, 2024

    Shilin Yan, Ouxiang Li, Jiayin Cai, Yanbin Hao, Xi- aolong Jiang, Yao Hu, and Weidi Xie. A sanity check for ai-generated image detection.arXiv preprint arXiv:2406.19435, 2024. 7, 8

    Pith/arXiv arXiv 2024

  39. [39]

    Orthogonal subspace decom- position for generalizable AI-generated image detection

    Zhiyuan Yan, Jiangming Wang, Peng Jin, Ke-Yue Zhang, Chengchun Liu, Shen Chen, Taiping Yao, Shouhong Ding, Baoyuan Wu, and Li Yuan. Orthogonal subspace decom- position for generalizable AI-generated image detection. In Forty-second International Conference on Machine Learn- ing, 2025. 2

    2025

  40. [40]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025. 6

    Pith/arXiv arXiv 2025

  41. [41]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InThe eleventh international conference on learning representations, 2022. 3

    2022

  42. [42]

    Towards universal image embed- dings: A large-scale dataset and challenge for generic image representations

    Nikolaos-Antonios Ypsilantis, Kaifeng Chen, Bingyi Cao, M´ario Lipovsk `y, Pelin Dogan-Sch ¨onberger, Grzegorz Makosa, Boris Bluntschli, Mojtaba Seyedhosseini, Ond ˇrej Chum, and Andr´e Araujo. Towards universal image embed- dings: A large-scale dataset and challenge for generic image representations. InProceedings of the ieee/cvf international conference...

    2023

  43. [43]

    Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers

    Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. InIcml, 2001. 4

    2001

  44. [44]

    Ivy-fake: A unified explainable framework and benchmark for image and video aigc detection.arXiv preprint arXiv:2506.00979,

    Wayne Zhang, Changjiang Jiang, Zhonghao Zhang, Chenyang Si, Fengchang Yu, and Wei Peng. Ivy-fake: A unified explainable framework and benchmark for image and video aigc detection.arXiv preprint arXiv:2506.00979,

    Pith/arXiv arXiv

  45. [45]

    Breaking semantic arti- facts for generalized ai-generated image detection.Advances in Neural Information Processing Systems, 37:59570–59596,

    Chende Zheng, Chenhao Lin, Zhengyu Zhao, Hang Wang, Xu Guo, Shuai Liu, and Chao Shen. Breaking semantic arti- facts for generalized ai-generated image detection.Advances in Neural Information Processing Systems, 37:59570–59596,

  46. [46]

    Patchcraft: Exploring texture patch for efficient ai-generated image detection.arXiv preprint arXiv:2311.12397, 2023

    Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Patchcraft: Exploring texture patch for efficient ai-generated image detection.arXiv preprint arXiv:2311.12397, 2023. 7, 8

    Pith/arXiv arXiv 2023

  47. [47]

    Breaking latent prior bias in detectors for generaliz- able aigc image detection.arXiv preprint arXiv:2506.00874,

    Yue Zhou, Xinan He, KaiQing Lin, Bin Fan, Feng Ding, and Bin Li. Breaking latent prior bias in detectors for generaliz- able aigc image detection.arXiv preprint arXiv:2506.00874,

    Pith/arXiv arXiv

  48. [48]

    Aigi-holmes: Towards explainable and gener- alizable ai-generated image detection via multimodal large language models.arXiv preprint arXiv:2507.02664, 2025

    Ziyin Zhou, Yunpeng Luo, Yuanchen Wu, Ke Sun, Jiayi Ji, Ke Yan, Shouhong Ding, Xiaoshuai Sun, Yunsheng Wu, and Rongrong Ji. Aigi-holmes: Towards explainable and gener- alizable ai-generated image detection via multimodal large language models.arXiv preprint arXiv:2507.02664, 2025. 2, 3

    Pith/arXiv arXiv 2025

  49. [49]

    All” denotes the full evaluation setting under this protocol, and “Overall

    Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: a million-scale benchmark for de- tecting ai-generated image. InProceedings of the 37th Inter- national Conference on Neural Information Processing Sys- tems, Red Hook, NY , USA, 2023. Curran Associates Inc. 7, 8 11 AgentFoX: ...

    arXiv 2023