pith. sign in

arxiv: 2605.25009 · v1 · pith:OQKWTFLVnew · submitted 2026-05-24 · 💻 cs.CV

ClueAegis: Heuristic-to-Reasoning Cognitive-skill Learning for Unified Evidence-based Synthetic Image Detection

Huangsen Cao , Hongkang Chu , Yuxi Li , Ying Zhang , Chen Li , Jing Lyu , Yongwei Wang , Yu Zhao
show 1 more author
Fei Wu
This is my paper

Pith reviewed 2026-06-30 12:29 UTC · model grok-4.3

classification 💻 cs.CV
keywords synthetic image detectionforensic reasoningcognitive skillsexplainable detectionheuristic cluesagentic frameworkcross-domain generalizationevidence-based analysis
0
0 comments X

The pith

A two-stage agentic system extracts perceptual clues, selects forensic skills, and reasons over evidence to detect synthetic images more robustly than end-to-end classifiers.

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

The paper claims that synthetic image detection improves when reframed as a cognitive process that first pulls heuristic visual clues from an image, then picks the right forensic skill, and finally runs skill-specific reasoning to gather evidence and decide. This matters because current monolithic detectors struggle with new generators and offer no insight into their verdicts. The authors support the approach with a new benchmark that annotates images for distinct cognitive skills and with a framework that chains the stages into transparent trajectories. Experiments show the method reaches state-of-the-art accuracy while generalizing better across domains and remaining more interpretable.

Core claim

ClueAegis reformulates synthetic image detection as a configurable multi-skill reasoning process: given an input image the framework extracts heuristic perceptual clues, selects the optimal forensic skill, and performs skill-conditioned reasoning through toolchains for evidence extraction and decision making, thereby bridging perception, skill selection, and forensic reasoning.

What carries the argument

The two-stage agentic framework that performs heuristic skill selection followed by evidence-guided reasoning through skill-conditioned toolchains.

If this is right

  • The method attains state-of-the-art detection accuracy on existing benchmarks.
  • Cross-domain generalization and robustness both increase relative to monolithic detectors.
  • Reasoning trajectories become transparent and forensic evidence is produced in structured form.
  • The system supplies a more explainable alternative to conventional end-to-end classifiers.

Where Pith is reading between the lines

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

  • The modular skill-selection design could transfer to related tasks such as video deepfake detection or image manipulation localization.
  • Because each skill is explicitly chosen and documented, the framework may reduce the amount of labeled data needed for new generator families.
  • The same heuristic-to-reasoning pattern might improve other perceptual judgment problems that currently rely on opaque neural classifiers.

Load-bearing premise

Breaking synthetic image detection into selectable heuristic perceptual clues and discrete forensic cognitive skills yields better evidence extraction and decisions than treating the task as a single end-to-end classification problem.

What would settle it

A controlled test in which ClueAegis records lower accuracy or weaker cross-generator robustness than a standard end-to-end classifier on a fresh collection of synthetic images produced by an unseen generative model.

Figures

Figures reproduced from arXiv: 2605.25009 by Chen Li, Fei Wu, Hongkang Chu, Huangsen Cao, Jing Lyu, Ying Zhang, Yongwei Wang, Yuxi Li, Yu Zhao.

Figure 1
Figure 1. Figure 1: (a) Prior synthetic image detection methods; [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ClueAegis-Bench. (a) Images are mapped from an entangled real/fake space into skill￾specific classification subspaces. (b) Skill-wise parti￾tions across heterogeneous sources form ClueAegis￾Bench with diverse synthetic distributions. 3 ClueAegis-Bench In cognitive psychology, dual-process theory dis￾tinguishes System 1—heuristic, intuitive percep￾tion—and System 2—deliberate analytical reason￾i… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of ClueAegis. A two-stage pipeline consisting of System 1 (heuristic) and System 2 (reasoning). tary types of authenticity cues, including lighting consistency (Light), shadow consistency (Shad), physical consistency (Phys), common-sense rea￾soning (CS), functional inconsistency (Func), OCR consistency (OCR), human anatomy (Human), re￾gion analysis (Region), animal anatomy (Animal), frequency cons… view at source ↗
Figure 4
Figure 4. Figure 4: Performance with different backbones. Effect of Different Backbone Models. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results on additional benchmarks. nityAI (Li et al., 2026)) and compare it with sev￾eral state-of-the-art methods. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: presents the visualization of ClueAegis￾Bench. The benchmark organizes images into skill-driven forensic subspaces according to distinct cognitive reasoning cues, covering diverse synthetic image distributions and visual contents. Each subspace corresponds to a specific forensic skill and is associated with structured reasoning annotations, enabling skill-aware evaluation beyond conventional real-versus-fa… view at source ↗
Figure 8
Figure 8. Figure 8: Training dynamics of our model. (a) Training [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance across different visual content [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Prompt template for Lighting Consistency. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Lighting Consistency. Skill-based forensic reasoning based on illumination consistency analysis. External lighting analysis models are further incorporated to assist the examination of spatial light directions and shadow relationships. G Qualitative Results In this section, we present qualitative examples of ClueAegis under all 12 predefined forensic skills. Specifically, we visualize the corresponding re… view at source ↗
Figure 11
Figure 11. Figure 11: Shadow Consistency. Skill-based forensic reasoning based on shadow consistency analysis. External shadow extraction and object segmentation models are further incorporated to examine the spatial coherence among objects, illumination, and projected shadows. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Physical Consistency. Skill-based forensic reasoning from a physical perspective, analyzing whether image content conforms to real-world physical laws and spatial interaction constraints. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Common Sense. Skill-based forensic reasoning based on commonsense consistency analysis, evaluating whether scene content, object relationships, and semantic interactions conform to real-world expectations and everyday human cognition. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Functional Inconsistency. Skill-based forensic reasoning based on functional consistency analysis, eval￾uating whether objects and scene components conform to realistic functionalities and real-world usage relationships. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: OCR Consistency. Skill-based forensic reasoning based on textual consistency analysis. External OCR models are further incorporated to examine recognized text content, typography, and semantic alignment consistency. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Human Anatomy. Skill-based forensic analysis that evaluates whether human body structures, anatomi￾cal proportions, and articulation relationships conform to realistic human anatomy principles. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Region Analysis. Skill-based forensic analysis focusing on localized scene regions and object-level details to identify potential inconsistencies in local structures, textures, and semantic compositions. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Frequency Consistency. Skill-based forensic analysis that incorporates frequency-spectrum analysis models to examine whether the image exhibits realistic frequency distributions and spectral consistency patterns. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Frequency Consistency. Skill-based forensic analysis that incorporates frequency-spectrum analysis models to examine whether the image exhibits realistic frequency distributions and spectral consistency patterns. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Pixel Consistency. Skill-based forensic analysis that leverages pixel-level forensic models to examine local pixel distributions, texture continuity, and low-level visual inconsistencies. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Transformation Consistency. Skill-based forensic analysis that evaluates robustness under geometric rotation and color transformations, examining whether structural layout and chromatic relationships remain consis￾tent and physically plausible after such changes. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_21.png] view at source ↗
read the original abstract

The rapid advancement of generative models has made synthetic images increasingly realistic, challenging reliable detection. Existing methods are often limited to end-to-end classification or monolithic reasoning, and thus fail to model structured forensic reasoning and heterogeneous visual evidence. We revisit synthetic image detection from a cognitive perspective and propose a \textit{Heuristic-to-Reasoning} cognitive skill learning framework for evidence-based forensic analysis. Given an input image, our framework first extracts heuristic perceptual clues, selects the optimal forensic skill, and then performs skill-conditioned reasoning for evidence extraction and decision making. To support this paradigm, we introduce \textbf{ClueAegis-Bench}, which decomposes synthetic image detection into explicitly annotated forensic cognitive skills for structured evaluation beyond binary classification. Based on this benchmark, we propose \textbf{ClueAegis} (\underline{C}ognitive-skill \underline{L}earning for \underline{U}nified \underline{E}vidence-based Synthetic Image Detection), a two-stage agentic framework that conducts heuristic skill selection followed by evidence-guided reasoning through skill-conditioned toolchains. This design reformulates synthetic image detection as a configurable multi-skill reasoning process that bridges perception, skill selection, and forensic reasoning. Extensive experiments show that ClueAegis achieves state-of-the-art performance while improving cross-domain generalization and robustness. It also provides transparent reasoning trajectories and structured forensic evidence, offering a more explainable alternative to conventional end-to-end detectors.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes ClueAegis, a two-stage agentic framework for synthetic image detection. Given an input image, it extracts heuristic perceptual clues, selects an optimal forensic skill, and performs skill-conditioned reasoning for evidence extraction and decision making. It introduces ClueAegis-Bench, which decomposes detection into explicitly annotated forensic cognitive skills. The framework is claimed to reformulate detection as a configurable multi-skill reasoning process that bridges perception, skill selection, and forensic reasoning, achieving SOTA performance with improved cross-domain generalization, robustness, and explainability over end-to-end detectors.

Significance. If the performance and generalization claims hold with proper validation, the work could advance synthetic image detection by moving beyond monolithic classification to structured, skill-based forensic reasoning, potentially yielding more robust and interpretable systems. The introduction of ClueAegis-Bench for cognitive-skill evaluation is a constructive contribution that could support future research on evidence-based detection.

major comments (2)
  1. [Abstract] Abstract: The central claims of SOTA performance, improved cross-domain generalization, and robustness are asserted without any quantitative results, baselines, dataset sizes, error bars, or experimental protocol. This absence is load-bearing because the abstract itself states that 'extensive experiments show' these outcomes, yet supplies no supporting data.
  2. [Abstract] Abstract: No ablations, component-wise comparisons, or controls are referenced that isolate the contribution of the heuristic skill selection and two-stage reasoning versus a direct clue-to-decision mapping. This leaves the key premise—that the explicit heuristic-to-skill decomposition itself drives the claimed gains—unverified and untested against monolithic alternatives.
minor comments (1)
  1. [Abstract] Abstract: The acronym expansion for ClueAegis is given with underlines as Cognitive-skill Learning for Unified Evidence-based Synthetic Image Detection; verify that the full manuscript maintains consistent acronym usage and definition.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for these focused comments on the abstract. We agree that the abstract should better ground its claims and will revise accordingly while preserving conciseness.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claims of SOTA performance, improved cross-domain generalization, and robustness are asserted without any quantitative results, baselines, dataset sizes, error bars, or experimental protocol. This absence is load-bearing because the abstract itself states that 'extensive experiments show' these outcomes, yet supplies no supporting data.

    Authors: We accept this observation. The current abstract is high-level by design, but the full manuscript contains the requested details (ClueAegis-Bench statistics, baseline comparisons, cross-domain splits, and error bars). In revision we will insert one or two concise quantitative highlights (e.g., peak accuracy and generalization gap) and a brief protocol reference to make the claims immediately verifiable from the abstract itself. revision: yes

  2. Referee: [Abstract] Abstract: No ablations, component-wise comparisons, or controls are referenced that isolate the contribution of the heuristic skill selection and two-stage reasoning versus a direct clue-to-decision mapping. This leaves the key premise—that the explicit heuristic-to-skill decomposition itself drives the claimed gains—unverified and untested against monolithic alternatives.

    Authors: The manuscript does contain the requested ablations and controls in the experimental section, directly comparing the two-stage heuristic-to-reasoning pipeline against direct clue-to-decision and end-to-end baselines. Because the abstract is a summary, it does not enumerate every ablation. We will add a short clause noting that component ablations confirm the contribution of skill selection and conditioned reasoning, thereby linking the abstract claim to the supporting evidence already present in the paper. revision: partial

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The abstract and description introduce a new two-stage agentic framework and benchmark without any equations, fitted parameters, self-citations, or derivations that reduce claims to inputs by construction. The central premise of heuristic-to-reasoning decomposition is presented as a design choice supported by experimental results rather than a self-referential reduction. No load-bearing steps match the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework depends on the domain assumption that synthetic image detection can be usefully decomposed into discrete, selectable forensic cognitive skills whose application yields structured evidence superior to end-to-end classification.

axioms (1)
  • domain assumption Synthetic image detection benefits from explicit decomposition into heuristic perceptual clues and selectable forensic cognitive skills
    This decomposition is the stated foundation for the two-stage agentic framework described in the abstract.
invented entities (1)
  • ClueAegis-Bench no independent evidence
    purpose: Decomposes synthetic image detection into explicitly annotated forensic cognitive skills for structured evaluation
    New benchmark introduced to support the cognitive-skill paradigm.

pith-pipeline@v0.9.1-grok · 5815 in / 1344 out tokens · 29702 ms · 2026-06-30T12:29:14.614622+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

14 extracted references · 7 canonical work pages · 5 internal anchors

  1. [1]

    Qwen3-VL Technical Report

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631. Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky, and Roy Schwartz. 2023. Breaking common sense: Whoops! a vision-and-language benchmark of syn- thetic and compositional images. InProceedings of the IEEE/CVF International Conference on Com- puter ...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    InEuropean con- ference on computer vision, pages 103–120

    What makes fake images detectable? under- standing properties that generalize. InEuropean con- ference on computer vision, pages 103–120. Springer. Ruoxin Chen, Jiahui Gao, Kaiqing Lin, Keyue Zhang, Yandan Zhao, Isabel Guan, Taiping Yao, and Shouhong Ding. 2025a. Task-model alignment: A simple path to generalizable ai-generated image de- tection.arXiv pre...

    work page arXiv 2025

  3. [3]

    Ivy-Fake: A Unified Explainable Framework and Benchmark for Image and Video AIGC Detection

    Ivy-fake: A unified explainable framework and benchmark for image and video aigc detection. arXiv preprint arXiv:2506.00979. Daniel Kahneman. 2011. Thinking, fast and slow.Far- rar, Straus and Giroux. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, and 1 ...

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  4. [4]

    InICASSP 2019- 2019 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 2307–

    Capsule-forensics: Using capsule networks to detect forged images and videos. InICASSP 2019- 2019 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 2307–

    2019

  5. [5]

    Utkarsh Ojha, Yuheng Li, and Yong Jae Lee

    IEEE. Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. 2023. To- wards universal fake image detectors that general- ize across generative models. InProceedings of the IEEE/CVF conference on computer vision and pat- tern recognition, pages 24480–24489. Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. 2023. Deeplsd: Line segment d...

    2023

  6. [6]

    Shadows don’t lie and lines can’t bend! gen- erative models don’t know projective geometry... for now. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 28140–28149. 10 Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. 2023. Hugging- gpt: Solving ai tasks with chatgpt and its friend...

    work page arXiv 2023

  7. [7]

    Qwen-Image Technical Report

    Dire for diffusion-generated image detection. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 22445–22455. Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, and 1 others. 2025. Qwen-image technical report.arXiv preprint arXiv:2508.02324. Chenfei Wu, Sh...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  8. [8]

    Visual ChatGPT: Talking, Drawing and Editing with Visual Foundation Models

    Visual chatgpt: Talking, drawing and edit- ing with visual foundation models.arXiv preprint arXiv:2303.04671. Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, and 1 others. 2024. Au- togen: Enabling next-gen llm applications via multi- agent conversations. InFirst conference on lang...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning

    Skillrl: Evolving agents via recursive skill- augmented reinforcement learning.arXiv preprint arXiv:2602.08234. Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. 2025. Fakeshield: Ex- plainable image forgery detection and localization via multi-modal large language models. InInter- national Conference on Learning Representations...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Use Image 1 as the reference for scene structure and illumination context

  11. [11]

    Compare Image 2 with Image 1 for lighting direction, intensity, highlights, and reflections

  12. [12]

    Check whether visible shadows and geometric cues are consistent with a coherent light source

  13. [13]

    Identify unrealistic lighting patterns, abrupt illumination changes, or physically implausible light behavior

  14. [14]

    transform_params

    Decide whether lighting inconsistencies provide evidence of synthetic generation. Figure 9: Prompt template for Lighting Consistency. The prompt includes input images, auxiliary DeepLSD- based lighting analysis, and a structured reasoning checklist for real-vs-synthetic classification. 17 Skill 1: Lighting Consistency (example1) FAKE HEURISTIC <think>The ...