Paper deep dive

FACT: A Forensic Agent with Compiled Tool-Use Trajectories for AI-Generated Image Detection

Jiaoyang Chen, Bin Hu, Jingyu Hu, Kun Zhou, Qin Zhang, Zhengzhe Liu

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 51

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%

Last extracted: 9/9/2026, 5:11:24 AM

Summary

The paper introduces FACT (Forensic Agent with Compiled Tool-use Trajectories), a novel framework for AI-generated image detection that treats detection as an open-world, image-conditioned tool-use problem. Unlike fixed detectors, FACT learns a policy to selectively call forensic tools (model-based experts and signal-level probes), interpret evidence, and stop when sufficient evidence is collected. The method employs an Evolve-Distill-Refine pipeline: evolving a forensic skill via execution-verified loops, distilling these trajectories into a compact agent via supervised fine-tuning, and refining the policy with cost-aware GRPO to optimize for accuracy and efficiency. FACT achieves state-of-the-art performance across multiple public and internal benchmarks, including on unseen generators and deepfakes.

Entities (15)

FACT · modelframework · 100%Evolve-Distill-Refine · methodology · 95%AIDE · model · 90%AIGCDetectionBenchmark · benchmark · 90%Chameleon · benchmark · 90%Community Forensics · modeldataset · 90%GRPO · algorithm · 90%HydraFake · benchmark · 90%LOKI · benchmark · 90%NewGen-900 · benchmark · 90%SDAIE · model · 90%SIDA · model · 90%STD3K · benchmark · 90%Veritas · model · 90%UniShield · model · 85%

Relation Signals (16)

FACT → achievesbestscoreon → STD3K

confidence 95% · the final refined FACT checkpoint achieves 91.40% [on STD3K]

FACT → achievesbestscoreon → AIGCDetectionBenchmark

confidence 95% · FACT obtains 98.77% on AIGCDetectionBenchmark

FACT → achievesbestscoreon → HydraFake

confidence 95% · FACT obtains 94.44% on HydraFake

FACT → achievesbestscoreon → Chameleon

confidence 95% · FACT obtains 87.53% on Chameleon

FACT → achievesbestscoreon → LOKI

confidence 95% · FACT obtains 88.35% balanced accuracy on LOKI

FACT → achievesbestscoreon → NewGen-900

confidence 95% · FACT obtains 82.44% [on NewGen-900]

FACT → usesmethodology → Evolve-Distill-Refine

confidence 95% · FACT follows an Evolve–Distill–Refine pipeline

FACT → outperforms → Community Forensics

confidence 90% · FACT achieves the best performance among all compared methods, including... Community Forensics

FACT → outperforms → SDAIE

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:AI-generated image detection is increasingly open-world: new image generators produce highly realistic images that make visual artifacts harder to identify. Existing detectors usually rely on a fixed set of forensic cues, so a detector that works well for one generator family may fail on another. We introduce FACT (Forensic Agent with Compiled Tool-use Trajectories), which learns an image-conditioned tool-use policy for forensic analysis. Instead of applying a fixed detector, FACT decides which forensic tools to call, interprets the returned evidence, and stops when sufficient evidence has been collected. FACT follows an Evolve--Distill--Refine pipeline: it evolves an execution-verified forensic skill, compiles the skill into action--observation tool-use trajectories, distills them into a compact agent, and refines the policy with cost-aware GRPO. Across two internal and four public benchmarks, FACT achieves the best performance among all compared methods, including on recent unseen generators, deepfakes, and manipulated images.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

48,554 characters extracted from source content.

Expand or collapse full text

FACT: A Forensic Agent with Compiled Tool-Use Trajectories for AI-Generated Image Detection Jiaoyang Chen Bin Hu Jingyu Hu Kun Zhou Qin Zhang Zhengzhe Liu Abstract AI-generated image detection is increasingly open-world: new image generators produce highly realistic images that make visual artifacts harder to identify. Existing detectors usually rely on a fixed set of forensic cues, so a detector that works well for one generator family may fail on another. We introduce FACT (Forensic Agent with Compiled Tool-use Trajectories), which learns an image-conditioned tool-use policy for forensic analysis. Instead of applying a fixed detector, FACT decides which forensic tools to call, interprets the returned evidence, and stops when sufficient evidence has been collected. FACT follows an Evolve–Distill–Refine pipeline: it evolves an execution-verified forensic skill, compiles the skill into action–observation tool-use trajectories, distills them into a compact agent, and refines the policy with cost-aware GRPO. Across two internal and four public benchmarks, FACT achieves the best performance among all compared methods, including on recent generators held out from all FACT training stages, deepfakes, and manipulated images. 1Lingnan University, Hong Kong 2Harbin Institute of Technology, Shenzhen 3The University of Hong Kong 4The Chinese University of Hong Kong 5Shenzhen University *Work done during an internship at Lingnan University, Hong Kong. zhengzheliu@LN.edu.hk Introduction The visual boundary between photographs and AI-generated images is rapidly disappearing. Modern generators reproduce coherent geometry, realistic materials, legible text, and plausible lighting, leaving few artifacts that casual inspection can reliably identify. Figure 2 turns this ambiguity into a simple question: which images are synthetic? The same ambiguity also challenges automatic detection. As image generators become more realistic and diverse, a cue that separates real and synthetic images for one generator family may not remain reliable for another. Image authentication is therefore an open-world problem. This open-world setting is difficult for a single fixed detector because existing methods rely on different forensic priors: CNN-synthesis artifacts (Wang et al. 2020); representations from large pretrained vision–language models (Ojha et al. 2023); generator-diverse training across thousands of models in Community Forensics (Park and Owens 2024); high-level semantic and high-/low-frequency features in AIDE (Yan et al. 2025a); camera-derived photographic priors in SDAIE (Zhong et al. 2025); multimodal detection, localization, and explanation in SIDA (Huang et al. 2024); and pattern-aware planning and self-reflection in Veritas (Tan et al. 2025). These cues are useful, but their reliability varies across generator families. In our balanced evaluation in Table 1, Veritas, AIDE, Community Forensics, and SDAIE each lead on different generator subsets. A fixed detector therefore commits to one evidence profile before seeing the image. The central question is instead: which evidence should be acquired for this particular image? Figure 1: Can you tell which images are real and which are generated by AI?22 2 In Fig. 2, the first two images are AI-generated, while the last two are real photographs. Method Nano Banana Nano Pro GPT- Image-1 SD3.5 L SD3 M Z-Image Turbo VERITAS 0.74 0.75 0.76 0.50 0.64 0.35 SIDA 0.70 0.61 0.51 0.53 0.58 0.57 SDAIE 0.64 0.54 0.36 0.69 0.70 1.00 AIDE 0.52 0.84 0.55 0.49 0.50 0.98 Comm. 0.67 0.56 0.59 0.52 0.83 0.71 Table 1: Detector specialization across recent generators. Each column is a balanced real-versus-synthetic subset. “Comm.” denotes Community Forensics; L and M denote Large and Medium. Darker cells indicate higher accuracy, and bold values mark the best detector in each subset. We address this image-conditioned evidence-acquisition problem with an agentic formulation: the model must decide which evidence to acquire, interpret returned observations, and stop when the evidence is sufficient. The most recent published agentic framework, UniShield, takes an initial step by routing each image to a predicted forgery domain and expert type before executing the selected detector (Huang et al. 2025). However, this remains closer to one-step routing than to a full forensic investigation. In open-world AI-generated image detection, useful evidence can vary from image to image even within the same broad forgery category. A forensic agent should therefore update its decision as new observations arrive and determine whether additional evidence is still needed. The limitation of one-step routing suggests that a forensic agent must do more than choose a detector once. A deployable agent must satisfy three requirements. First, it needs procedural knowledge: image-level labels do not specify how to investigate, which evidence to acquire, or how to reconcile conflicting cues. Second, it needs specialization: a general-purpose model can discover procedures, but deployment requires a compact agent specialized for forensic tool use and evidence-conditioned decisions. Third, it needs budget-aware stopping: the agent must decide when another tool call justifies its cost and when further evidence may become redundant or misleading under changing image quality and disagreement among heterogeneous forensic experts. In this paper, we introduce FACT (Forensic Agent with Compiled Tool-use Trajectories), an Evolve–Distill–Refine pipeline for learning a deployable forensic investigation policy. To acquire procedural knowledge, Evolve represents forensic expertise as an explicit skill and improves it through a propose–execute–verify loop, accepting a revision only when rerunning the full tool interaction improves measured behavior. To achieve specialization, Distill uses the verified skill to guide a strong teacher in the real tool environment, compiling the procedure into action–observation trajectories and behaviorally cloning them into a compact agent. To enable budget-aware stopping, Refine applies cost-aware GRPO after distillation, rewarding final correctness while penalizing unnecessary evidence acquisition. At deployment, FACT uses only the compact forensic agent. For each test image, the agent decides whether to call a forensic tool, use the returned observation, or stop with a final real/AI decision. The skill-evolution components and the general model are used only during training. We evaluate FACT on two balanced internal benchmarks and four public benchmarks. STD3K evaluates robustness under source diversity, containing 3,200 images from 21 generator/source groups and multiple real-image sources; the final refined FACT checkpoint achieves 91.40%. NewGen-900 evaluates generalization to nine recent generator families held out from all FACT training stages; FACT obtains 82.44% without generator-specific fine-tuning. On public benchmarks, FACT obtains 87.53% on Chameleon (Yan et al. 2025a), 88.35% balanced accuracy on LOKI (Ye et al. 2025), 98.77% on AIGCDetectionBenchmark (Zhong et al. 2023), and 94.44% on HydraFake (Tan et al. 2025). Across all six benchmarks, FACT achieves the best performance among the methods compared under each corresponding protocol, covering visually challenging synthetic images, recent generators held out from all FACT training stages, and broader forgery settings involving deepfakes and manipulated images. Our contributions are threefold: • We formulate AI-generated image detection as an image-conditioned tool-use problem. Rather than applying a fixed detector, the agent selectively acquires signal-level and model-based forensic evidence, updates its decision from returned observations, and stops when the evidence is sufficient. • We propose an Evolve–Distill–Refine pipeline that learns such a forensic agent from image-level labels. It evolves an executable forensic skill, compiles the skill into action–observation tool-use trajectories, distills them into a specialized policy, and further optimizes tool selection and stopping with cost-aware GRPO. • We demonstrate strong performance and broad applicability across six benchmarks. FACT consistently outperforms its constituent tools, majority-vote ensembling, and recent detector baselines, while generalizing to recent generators held out from all FACT training stages and remaining effective on deepfakes and manipulated images. Related Work AI image generation. Image synthesis has evolved from GAN-based generation (Goodfellow et al. 2014; Karras et al. 2019), through diffusion and latent-diffusion models (Ho et al. 2020; Rombach et al. 2022), to transformer- and flow-based scaling (Peebles and Xie 2023; Esser et al. 2024). Recent systems further unify generation, multimodal understanding, and interactive editing. GPT-Image, GPT-Image-2, Nano Banana and Nano Banana Pro, Qwen-Image, and OmniGen2 support capabilities such as multi-image conditioning, conversational editing, text rendering, and modification of real photographs (OpenAI 2025; OpenAI 2026; Google 2025a; Google 2025b; Wu et al. 2025a; Wu et al. 2025b). These capabilities make synthetic images increasingly realistic, controllable, and easy to revise, increasing authenticity risks for image verification. AI-generated image detection. AI-generated image detection has evolved from recognizing generator-specific artifacts to learning more general and semantically informed notions of authenticity. Early methods exploit GAN fingerprints, upsampling artifacts, global texture inconsistency, and local frequency or texture irregularities (Wang et al. 2020; Liu et al. 2020); these cues can be highly effective on represented generators, but often weaken under unseen architectures, compression, resizing, and editing. Later methods improve cross-generator robustness by transferring pretrained visual representations (Ojha et al. 2023), using reconstruction or diffusion-based discrepancies (Wang et al. 2023), or expanding training diversity across many generators, as in Community Forensics (Park and Owens 2024). Recent detectors further introduce heterogeneous forensic priors: SDAIE learns a photographic prior from camera metadata using only real camera images (Zhong et al. 2025); AIDE combines semantic representations with high- and low-frequency evidence (Yan et al. 2025a); SIDA unifies classification and textual explanation in a multimodal model (Huang et al. 2024); and Veritas incorporates planning and self-reflection for pattern-aware reasoning (Tan et al. 2025). These methods improve generalization, but their strengths remain generator- and evidence-dependent: a cue that is reliable for one generator family may be weak or misleading for another. FACT therefore does not introduce another fixed forensic prior. Instead, it treats existing priors as frozen callable experts and learns an image-conditioned policy for deciding which evidence should be acquired before making a decision. Self-improving agents and agentic forensics. Tool-using agents interleave reasoning with external actions, and recent self-improving agents learn from reflection, interaction feedback, trajectories, or reusable skills (Yao et al. 2023; Shinn et al. 2023; Madaan et al. 2023; Zhao et al. 2024; Zheng et al. 2025; Chen et al. 2025; Yang et al. 2026). In image forensics, UniShield is the closest published agentic framework: it predicts a broad forgery category, selects a corresponding detector, and generates a report (Huang et al. 2025). FACT differs by learning a full tool-use investigation procedure rather than a category-to-detector routing rule: it compiles an execution-verified forensic skill into real action–observation trajectories, distills them into a compact agent, and refines the policy for cost-aware stopping. Method Overview and Problem Formulation Given an image x, FACT aims to determine whether it is real or AI-generated. To support this decision, FACT first constructs a forensic toolbox U from two complementary evidence sources: frozen model-based experts and lightweight signal-level probes. The goal is to learn a policy πθ _θ that selectively uses this toolbox before predicting y^∈real,AI y∈\real,AI\. At step t, the policy observes the interaction history ht=(x,a1,o1,…,at−1,ot−1),h_t=(x,a_1,o_1,…,a_t-1,o_t-1), and either calls a tool at∈a_t to obtain observation oto_t, or stops and outputs a final verdict. The complete trajectory and its acquisition cost are τ=(x,a1,o1,…,aT,oT,y^),C(τ)=∑t:at∈c(at),τ=(x,a_1,o_1,…,a_T,o_T, y), C(τ)= _t:a_t c(a_t), where c⁡(at)c(a_t) is the normalized cost of tool ata_t. As shown in Fig. 2, FACT learns πθ _θ in three stages. Stage I uses a general AI model to analyze labeled images with the toolbox and iteratively improve a textual forensic skill S; candidate skill revisions are accepted only after rerunning the tool-use process and verifying that they improve accuracy. Stage I fixes the verified skill S⋆S and uses it to generate real action–observation trajectories in the same tool environment, which are then distilled into a compact policy by SFT. Stage I refines the distilled policy with cost-aware GRPO, rewarding correct final decisions while penalizing tool cost, malformed answers, and invalid calls. During inference, FACT uses only the trained policy and the toolbox: for each image, the policy decides whether to answer directly, call more evidence tools, or stop. Figure 2: Overview of FACT. FACT trains a forensic agent with an Evolve–Distill–Refine pipeline. Skill evolution learns an executable investigation procedure from judged cases; trajectory distillation converts the skill into supervised action–observation tool-use data; and cost-aware GRPO teaches the agent when additional evidence is worth its cost. At inference, the trained agent adaptively calls forensic tools or stops with a final real/AI decision. Forensic Toolbox Construction Before training the agent, we first examine how existing detectors behave on different generator families. As shown in Table 1, detectors that are strong on one generator can be weaker on another. This observation suggests that the goal should not be to choose a single detector as the universal solution. Instead, FACT organizes complementary detectors and signal analyses into a toolbox U, so that the agent can learn which evidence is useful for each image. Model-based experts. The first part of the toolbox consists of five frozen model-based experts. These experts capture different types of forensic evidence and provide complementary signals for AI-generated image detection. SDAIE models the photographic regularities of real camera images through EXIF-induced supervision (Zhong et al. 2025). Community Forensics emphasizes broad discriminative coverage by training across thousands of generators (Park and Owens 2024). AIDE combines semantic features with high- and low-frequency evidence (Yan et al. 2025a). SIDA brings multimodal reasoning and textual explanation into the detection process (Huang et al. 2024). Veritas uses pattern-aware planning and self-reflection to reason about synthetic artifacts (Tan et al. 2025). These experts therefore capture different cues, including camera-photo consistency, large-scale generator diversity, semantic plausibility, multimodal reasoning, and reasoning-based artifact analysis. Their strengths across generator subsets motivate using them as callable experts rather than choosing one as the fixed detector. Signal-level probes. The second part of the toolbox contains eight lightweight signal-level probes. These probes are included because low-level texture, frequency, noise, and compression statistics have long been useful for fake-image detection (Wang et al. 2020; Liu et al. 2020), and they provide evidence that is different from the final prediction of a neural detector. FFT and phase statistics describe spectral behavior; wavelet and Laplacian analyses measure multiscale texture; noise residuals and JPEG analysis capture residual and compression consistency; gradient statistics and local patch consistency describe local structural regularity. These probes do not replace model-based experts. Rather, they give the agent simple and interpretable measurements that can be combined with expert outputs when deciding whether an image is real or AI-generated. Stage I: Tool-Use Skill Learning With the toolbox U fixed, Stage I learns how a general model MGM_G should use it for forensic judgment. The goal is not only to decide whether an image is real or AI-generated, but also to learn a reusable investigation procedure: which tools to call, why they are useful, how to interpret their outputs, how to handle conflicting evidence, and when to stop. Following recent agent works that store reusable experience as skills or reflection-derived procedures (Zhao et al. 2024; Zheng et al. 2025; Yang et al. 2026), FACT represents this procedure as a textual forensic skill S. In our implementation, S is a structured instruction covering tool selection, evidence interpretation, conflict handling, stopping criteria, and final-answer format. Updating S lets MGM_G become more specialized for AI-generated image detection without finetuning its parameters. At update step r, FACT samples a labeled skill-update batch ℬr=(xi,yi)i=1mB_r=\(x_i,y_i)\_i=1^m. The current skill SrS_r is given to MGM_G, which analyzes each image with access to the toolbox U. During the investigation, MGM_G sees the image and the tool observations it requests, but not the ground-truth label. Each run is recorded as a trajectory τi=(xi,ai,1,oi,1,…,ai,Ti,oi,Ti,y^i), _i=(x_i,a_i,1,o_i,1,…,a_i,T_i,o_i,T_i, y_i), where y^i y_i is the final prediction. After the batch is judged, the labels are revealed and the errors are summarized as r=(xi,yi,y^i,τi)i=1m.D_r=\(x_i,y_i, y_i, _i)\_i=1^m. The model then uses the current skill and this error summary to propose a revised skill: S~r=UpdateSkillMG(Sr,r). S_r=UpdateSkill_M_G(S_r,D_r). FACT evaluates the proposed revision on a separately sampled labeled verification batch r=(xj(v),yj(v))j=1n,V_r=\(x_j^(v),y_j^(v))\_j=1^n, which is not used to construct rD_r or propose S~r S_r. Let (y^j(v)(S),τj(v)(S))=ℱMG(S,xj(v),)( y_j^(v)(S), _j^(v)(S))=F_M_G(S,x_j^(v);U) denote running MGM_G on verification image xj(v)x_j^(v) with skill S and toolbox U. The verification accuracy is (S;r)=1|r|∑(xj(v),yj(v))∈r[y^j(v)(S)=yj(v)].J(S;V_r)= 1|V_r| _(x_j^(v),y_j^(v)) _r1\! [ y_j^(v)(S)=y_j^(v) ]. The skill is updated by the following replay check: Sr+1=S~r,⁡(S~r,r)>⁡(Sr,r),Sr,otherwise.S_r+1= cases S_r,&J( S_r;V_r)>J(S_r;V_r),\\ S_r,&otherwise. cases Thus, a revised instruction is accepted only when it improves rerun tool-use performance on the separate verification batch. Beyond single-batch updates, FACT also summarizes recent error summaries and proposes broader skill revisions. For example, if the model repeatedly over-trusts frequency artifacts on compressed real images, the skill can be revised to check compression consistency before making a final decision. If the model repeatedly overlooks spatially inconsistent evidence, the skill can place greater emphasis on patch-level consistency checks. These broader updates are retained only after the same replay check. The output of Stage I is a verified forensic skill S⋆S , which provides the procedure used to generate training trajectories in Stage I. Representative skill revisions and the complete initial and evolved skill texts are provided in the supplementary material. Stage I: Distilling Tool-Use Trajectories Stage I produces a verified forensic skill S⋆S , but this skill is still a textual procedure used by a general model. For deployment, we need a compact model that can execute the procedure directly: it should know when to call a tool, how to react to the returned evidence, and when to stop. Stage I therefore converts the skill into supervised tool-use trajectories and distills them into a compact policy πθ _θ. For each labeled training image (xi,yi)(x_i,y_i), the general model MGM_G follows S⋆S and interacts with the toolbox U. It selects a tool, receives the tool output, and then decides the next action until it produces a final decision. We record this process as τi=(xi,ai,1,oi,1,…,ai,Ti,oi,Ti,yi), _i=(x_i,a_i,1,o_i,1,…,a_i,T_i,o_i,T_i,y_i), where ai,ta_i,t is a tool call or a stopping action, oi,to_i,t is the corresponding tool output, and yiy_i provides the supervised final answer. The key supervision is not only the final label, but the intermediate action sequence showing how evidence changes the next decision. We serialize each trajectory into a training sequence containing the image, task instruction, tool definitions, previous tool calls, returned tool outputs, and the next action to imitate. The compact policy is trained by supervised fine-tuning: ℒSFT=−1∑i,tmi,t∑i,tmi,tlogπθ(zi,t∣xi,zi,<t),L_SFT=- 1 _i,tm_i,t _i,tm_i,t _θ(z_i,t x_i,z_i,<t), where zi,tz_i,t is a token in the serialized trajectory and mi,t=1m_i,t=1 only for tokens that the agent should generate, such as tool calls, stopping decisions, and the final answer. Tokens from tool outputs are used as context but are not prediction targets. After this stage, πθ _θ becomes a compact forensic agent trained to imitate the tool-use behavior induced by S⋆S . It has learned the basic investigation procedure, while Stage I further optimizes how much evidence should be acquired under a tool-cost budget. Stage I: Cost-Aware Tool-Use Refinement The SFT policy has learned how to conduct a forensic investigation, but supervised imitation does not directly optimize the deployment objective: making a correct decision under a limited evidence budget. Stage I therefore refines the distilled policy with cost-aware GRPO. For each training image xix_i, we sample a group of K complete trajectories τikk=1K,τik=(xi,ai,1k,oi,1k,…,ai,Tikk,oi,Tikk,y^ik),\ _i^k\_k=1^K, _i^k=(x_i,a_i,1^k,o_i,1^k,…,a_i,T_i^k^k,o_i,T_i^k^k, y_i^k), from the current policy. Each trajectory is scored by a task–cost reward: R⁡(τik,yi)=Rtask(y^ik,yi)−λC(τik)−Rformat(τik)−Rinvalid(τik).R( _i^k,y_i)=R_task( y_i^k,y_i)-λ C( _i^k)-R_format( _i^k)-R_invalid( _i^k). The task term rewards the final decision, Rtask(y^,y)=+1,y^=y,−1,y^≠y,0,y^=uncertain,−1.2,no valid final answer,R_task( y,y)= cases+1,& y=y,\\ -1,& y≠ y,\\ 0,& y=uncertain,\\ -1.2,&no valid final answer, cases while the cost term penalizes cumulative tool acquisition, C(τ)=∑t:at∈c(at).C(τ)= _t:a_t c(a_t). The remaining terms penalize interaction failures: Rformat(τ) R_format(τ) =0.2 1[malformed final response], =0.2\,1[malformed final response], Rinvalid(τ) R_invalid(τ) =0.2Ninvalid(τ), =0.2\,N_invalid(τ), where Ninvalid(τ)N_invalid(τ) is the number of invalid tool calls. GRPO compares trajectories sampled for the same image. For trajectory τik _i^k, its group-relative advantage is Aik=R⁡(τik,yi)−1K∑ℓ=1KR⁡(τiℓ,yi)stdℓ(R⁡(τiℓ,yi))+ϵ.A_i^k= R( _i^k,y_i)- 1K _ =1^KR( _i ,y_i)std_ (R( _i ,y_i))+ε. Let zi,k,tz_i,k,t denote an agent-generated token in trajectory τik _i^k, including tool calls and final-answer tokens, and let mi,k,tm_i,k,t mask these trainable tokens. The clipped GRPO objective is gi,k,t=min⁡(ρi,k,tAik,clip⁡(ρi,k,t,1−ϵ,1+ϵ)Aik).g_i,k,t= \! ( _i,k,tA_i^k,\,clip( _i,k,t,1-ε,1+ε)A_i^k ). ℒGRPO=−1∑i,k,tmi,k,t∑i,k,tmi,k,tgi,k,t+βDKL(πθ∥πref).L_GRPO=- 1 _i,k,tm_i,k,t _i,k,tm_i,k,tg_i,k,t+β D_KL( _θ\| _ref). where ρi,k,t=πθ(zi,k,t∣xi,zi,k,<t)πold(zi,k,t∣xi,zi,k,<t), _i,k,t= _θ(z_i,k,t x_i,z_i,k,<t) _old(z_i,k,t x_i,z_i,k,<t), and πref _ref is the fixed SFT reference policy. This refinement does not ask RL to discover tool use from scratch. Instead, because the policy already imitates valid trajectories from Stage I, the reward mainly teaches whether another tool call improves the final decision enough to justify its cost. The coefficient λ controls this evidence budget: increasing λ encourages the agent to avoid unnecessary calls, while the task reward prevents it from stopping too early. Inference. At inference time, FACT uses the trained forensic agent πθ _θ to judge each test image. The agent predicts a tool-use sequence over the forensic toolbox U, where each step either calls a tool or terminates with Stop. Returned observations are added to the history and used to choose the next action. Once the agent stops, it outputs a real/AI verdict and brief rationale based on the evidence it acquired. The skill-learning and trajectory-synthesis components are used only during training. Experiments Experimental Setup Benchmarks. We evaluate FACT on two balanced internal benchmarks and four public benchmarks. The internal benchmarks are designed to isolate two deployment scenarios that are not fully covered by existing test sets. STD3K contains 3,200 images, with 1,600 generated images from 21 generator/source groups and 1,600 real images from LIVE, CLIVE, Flickr8K, and TID2013 (Sheikh et al. 2006; Ghadiyaram and Bovik 2016; Hodosh et al. 2013; Ponomarenko et al. 2015). It measures robustness under broad source diversity rather than performance on a single generator family. NewGen-900 contains 900 images, with 450 generated images and 450 real images from the same real-image sources. The fake images are drawn from nine recent generators: Nano Banana, Nano Banana Pro, GPT-Image-1, GPT-Image-2, SD3.5-Large, SD3-Medium, OmniGen2, Qwen-Image, and Z-Image Turbo. These generator families are held out from all FACT training stages, so NewGen-900 directly evaluates open-world cross-generator generalization. For public evaluation, we use benchmarks that emphasize different failure modes. Chameleon contains visually challenging AI-generated images that are often misclassified as real by existing detectors (Yan et al. 2025a); we evaluate on its official split. LOKI evaluates synthetic-data detection under a multimodal LMM protocol with real/synthetic judgments and explanation-oriented questions (Ye et al. 2025); we report balanced accuracy on its image-detection subset to match the public protocol. AIGCDetectionBenchmark covers a broad set of AI image generators, including 17 prevalent generative models (Zhong et al. 2023); we use its official image-detection test split, which also enables direct comparison with UniShield. HydraFake extends the evaluation beyond full-image generation to broader image-forgery settings, including face swapping, reenactment, attribute editing, personalization, relighting, restoration, and cross-domain deepfakes (Tan et al. 2025). We report the average accuracy across its official evaluation splits (Tan et al. 2025). Detailed dataset composition, sampling rules, and benchmark documentation are provided in the supplementary material. Baselines and metrics. We compare FACT with the five frozen model-based experts in the toolbox—SDAIE, Community Forensics, AIDE, SIDA, and Veritas—and their unweighted majority vote. For the public benchmarks, we additionally include published results under the corresponding benchmark protocols. Detailed baseline descriptions and the exact provenance of externally reported results are provided in the supplementary material. We report accuracy on STD3K, NewGen-900, Chameleon, and AIGCDetectionBenchmark; balanced accuracy on LOKI, BAcc=12(TPR+TNR)BAcc= 12(TPR+TNR); and average accuracy across the official HydraFake evaluation splits. Efficiency is measured by the average normalized acquisition cost C⁡(τ)C(τ). Implementation details. FACT uses GPT-5.1 for skill evolution and trajectory synthesis, and Qwen3-VL-8B-Instruct (Bai et al. 2025) as the compact forensic agent. Full training configurations, optimization settings, and tool-cost calibration are provided in the supplementary material. Comparison with Existing Works Tables 2–6 compare FACT with its constituent experts, majority voting, and representative published baselines on the internal and public benchmarks. Complete baseline results and their provenance are provided in the supplementary material. Unless otherwise stated, FACT denotes the Stage I refined checkpoint. Method STD3K NewGen-900 Community (Park and Owens 2024) 78.56 72.67 Veritas (Tan et al. 2025) 40.78 43.11 SIDA (Huang et al. 2024) 77.38 61.00 SDAIE (Zhong et al. 2025) 59.53 76.89 AIDE (Yan et al. 2025a) 44.31 73.44 Majority Vote 72.97 75.33 FACT 91.40 82.44 Table 2: Comparison on the internal benchmarks. All values are accuracies in percentages. All methods are evaluated using our fixed evaluation pipeline. On STD3K, the final refined FACT checkpoint reaches 91.40%, exceeding the strongest constituent expert by 12.84 percentage points and demonstrating robustness across diverse generator and real-image sources. On NewGen-900, the final refined FACT checkpoint obtains 82.44%, outperforming the strongest constituent expert by 5.55 points. All nine generator families are held out from skill evolution, trajectory distillation, and RL refinement, demonstrating open-world generalization without generator-specific fine-tuning. On Chameleon, FACT achieves 87.53%, exceeding DefakerOne, the strongest previously reported comparison, by 2.83 points. This indicates strong performance on visually challenging synthetic images. Method Accuracy NPR† (Tan et al. 2024) 57.81 AIDE† (Yan et al. 2025a) 65.77 Community (Park and Owens 2024) 78.44 Veritas (Tan et al. 2025) 58.23 SIDA (Huang et al. 2024) 66.50 SDAIE (Zhong et al. 2025) 58.86 Majority Vote 65.97 Ivy-xDetector‡ (Zhang et al. 2025) 73.17 DefakerOne‡ (GuangJian Team 2026) 84.70 FACT 87.53 Table 3: Comparison on Chameleon. † : Table 5 of (Yan et al. 2025a); ‡ : cited method papers. Unmarked results use our fixed pipeline. On LOKI, FACT achieves 88.35% balanced accuracy, surpassing EvoGuard by 1.97 percentage points under the reported public protocol. This shows that the learned policy remains effective under the multimodal evaluation setting. On AIGCDetectionBenchmark, FACT reaches 98.77%, exceeding the strongest comparison by 1.97 points. The gain over individual detectors and majority voting supports image-conditioned evidence acquisition over static aggregation. On HydraFake, FACT achieves 94.44% average accuracy, outperforming Veritas by 3.74 points and showing that FACT remains effective beyond fully generated images, including on deepfakes and edited images. Across all six benchmarks, FACT outperforms every constituent expert, majority voting, and the strongest available published comparison. Method BAcc. Effort† (Yan et al. 2025b) 74.09 FakeVLM† (Wen et al. 2025) 81.62 MIRROR† (Liu et al. 2026) 85.23 AIDE† (Yan et al. 2025a) 73.70 FakeShield† (Xu et al. 2025) 61.55 SIDA† (Huang et al. 2024) 61.92 FakeReasoning† (Gao et al. 2025) 64.78 Forensic-MoE† (Fang et al. 2025) 74.92 EvoGuard† (Zhu et al. 2026) 86.38 Community (Park and Owens 2024) 82.80 Veritas (Tan et al. 2025) 65.95 SDAIE (Zhong et al. 2025) 82.85 Majority Vote 83.20 FACT 88.35 Table 4: Comparison on LOKI. † : Table 2 of (Zhu et al. 2026). Unmarked results use our fixed pipeline. Method Accuracy AIDE† (Yan et al. 2025a) 92.80 FakeVLM† (Wen et al. 2025) 81.00 UniShield† (Huang et al. 2025) 94.20 Community (Park and Owens 2024) 96.80 Veritas (Tan et al. 2025) 69.60 SIDA (Huang et al. 2024) 59.90 SDAIE (Zhong et al. 2025) 96.00 ReAlign‡ (Huang et al. 2026) 96.14 Majority Vote 95.50 FACT 98.77 Table 5: Comparison on AIGCDetectionBenchmark. † : Table 6 of (Huang et al. 2025); ‡ : cited method paper. Unmarked results use our fixed pipeline. (a) Method Acc. FreqNet† 64.60 ProDet† 80.60 NPR† 69.80 AIDE† 70.60 Co-SPY† 84.70 D3† 81.10 Effort† 82.20 Qwen2.5-VL-7B† 54.10 InternVL3-8B† 58.30 MiMo-VL-7B† 72.50 GLM-4.1V-9B-T† 61.70 GPT-4o† 60.80 (b) Method Acc. Gemini-2.5-Pro† 78.90 M2F2-Det† 63.20 FakeShield† 60.80 SIDA-7B† 76.30 SIDA-13B† 69.80 FFAA† 64.00 FakeVLM† 77.30 Veritas† 90.70 AIDE (ours) 61.60 SDAIE (ours) 68.30 Community (ours) 76.30 FACT 94.44 Table 6: Comparison on HydraFake. Values are average accuracies in percentages. † : Table 1 of (Tan et al. 2025); unmarked results use our fixed pipeline. “T” abbreviates Think. The supplementary material additionally provides qualitative case analyses, full agent conversation traces, and the human evaluation. Ablation Studies (a) Training stages Var. S D R STD New Direct – – – 87.90 50.80 Skill ✓ – – 89.40 75.10 Distill. ✓ ✓ – 91.53 81.89 Refine ✓ ✓ ✓ 91.40 82.44 (b) Toolbox Sig. Exp. Acc. ✓ – 82.90 – ✓ 82.50 ✓ ✓ 90.70 Table 7: Pipeline and toolbox ablations. In (a), S, D, and R denote skill learning, trajectory distillation, and GRPO refinement; the variants are Direct, Skill-guided, Distilled, and Refined. In (b), Sig. and Exp. denote signal probes and model-based experts. Task Fmt. Valid Cost Acc. Acq. ✓ – – – 81.78 1.5300 ✓ ✓ – – 82.00 1.5700 ✓ ✓ ✓ – 82.56 1.5500 ✓ ✓ ✓ ✓ 82.44 1.3835 Table 8: Reward-component ablation on NewGen-900. Reward terms are added cumulatively; the full reward uses λ=0.01λ=0.01. Contribution of each training stage. Table 7(a) evaluates skill learning, trajectory distillation, and cost-aware refinement. Directly querying GPT-5.1 reaches 87.90% on STD3K but only 50.80% on NewGen-900. Adding the learned skill raises performance to 89.40% and 75.10%, showing that Stage I contributes procedural knowledge beyond direct visual judgment. Distilling skill-guided action–observation trajectories further improves performance to 91.53% and 81.89%. Cost-aware GRPO produces the final FACT checkpoint, changing accuracy to 91.40% on STD3K and 82.44% on NewGen-900. Thus, refinement trades a marginal 0.13-point decrease on STD3K for a 0.55-point gain on NewGen-900. In the reward-component experiment, the full cost-aware reward achieves 82.44% accuracy while reducing normalized acquisition cost from 1.5500 to 1.3835 (Table 8). Effect of toolbox composition. Table 7(b) compares signal probes, model-based experts, and their combination on STD3K + NewGen-900. The two families alone obtain 82.90% and 82.50%, while combining them reaches 90.70%. This confirms that low-level measurements and high-level detector outputs provide complementary evidence. Effect of reward components. Table 8 shows that format and valid-tool penalties improve NewGen-900 accuracy from 81.78% to 82.56%. Adding tool cost yields 82.44% accuracy and lowers acquisition cost from 1.5500 to 1.3835. Conclusion We presented FACT, a forensic agent that reframes AI-generated image detection as learning an investigation procedure rather than training another fixed detector. FACT converts image-level supervision into a tool-use skill, compiles it into action–observation trajectories, and distills them into a cost-aware agent. Across benchmarks, FACT improves over individual experts and static voting while generalizing to recent generators held out from all FACT training stages. The results show that evidence acquisition and stopping should be optimized jointly. The refined policy preserves detection performance while reducing normalized acquisition cost, avoiding redundant calls without abandoning difficult cases. FACT suggests that robustness may depend less on a universal detector and more on learning which evidence each image requires. Additional qualitative analyses and limitations are provided in the supplementary material. Upon publication, we will release the code, trained model, and benchmark data and metadata, subject to source licenses. References Bai et al. (2025) S. Bai, Y. Cai, R. Chen, et al. Qwen3-vl technical report. External Links: 2511.21631 Cited by: Implementation details.. Chen et al. (2025) Y. Chen, B. Xu, X. Wang, Y. Zhang, and Z. Mao Training llm-based agents with synthetic self-reflected trajectories and partial masking. External Links: 2505.20023 Cited by: Self-improving agents and agentic forensics.. Esser et al. (2024) P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, Z. English, K. Lacey, A. Goodwin, Y. Marek, and R. Rombach Scaling rectified flow transformers for high-resolution image synthesis. External Links: 2403.03206 Cited by: AI image generation.. Fang et al. (2025) M. Fang, Z. Li, L. Yu, Q. Yang, H. Xie, and Y. Zhang Forensic-moe: exploring comprehensive synthetic image detection traces with mixture of experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 17772–17782. Cited by: Table 4. Gao et al. (2025) Y. Gao, D. Chang, B. Yu, H. Qin, L. Chen, K. Liang, and Z. Ma FakeReasoning: towards generalizable forgery detection and reasoning. External Links: 2503.21210 Cited by: Table 4. Ghadiyaram and Bovik (2016) D. Ghadiyaram and A. C. Bovik Massive online crowdsourced study of subjective and objective picture quality. IEEE Transactions on Image Processing. Cited by: Benchmarks.. Goodfellow et al. (2014) I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio Generative adversarial nets. In NeurIPS, Cited by: AI image generation.. Google (2025a) Google Introducing gemini 2.5 flash image, our state-of-the-art image model. Note: https://developers.googleblog.com/en/introducing-gemini-2-5-flash-image/Accessed: 2026-06-30 Cited by: AI image generation.. Google (2025b) Google Introducing nano banana pro. Note: https://blog.google/innovation-and-ai/products/gemini-app/nano-banana-pro/Accessed: 2026-06-30 Cited by: AI image generation.. GuangJian Team (2026) GuangJian Team Venus-defakerone: unified fake image detection & localization. arXiv preprint arXiv:2605.14091. Cited by: Table 3. Ho et al. (2020) J. Ho, A. Jain, and P. Abbeel Denoising diffusion probabilistic models. In NeurIPS, Cited by: AI image generation.. Hodosh et al. (2013) M. Hodosh, P. Young, and J. Hockenmaier Framing image description as a ranking task: data, models and evaluation metrics. Journal of Artificial Intelligence Research. Cited by: Benchmarks.. Huang et al. (2026) Q. Huang, Z. Xu, X. Zhang, X. Yu, and J. Zhang ReAlign: generalizable image forgery detection via reasoning-aligned representation. arXiv preprint arXiv:2605.16080. Cited by: Table 5. Huang et al. (2025) Q. Huang, Z. Xu, X. Zhang, and J. Zhang UniShield: an adaptive multi-agent framework for unified forgery image detection and localization. External Links: 2510.03161 Cited by: Introduction, Self-improving agents and agentic forensics., Table 5, Table 5. Huang et al. (2024) Z. Huang, J. Hu, X. Li, Y. He, X. Zhao, B. Peng, B. Wu, X. Huang, and G. Cheng SIDA: social media image deepfake detection, localization and explanation with large multimodal model. External Links: 2412.04292 Cited by: Introduction, AI-generated image detection., Model-based experts., Table 2, Table 3, Table 4, Table 5. Karras et al. (2019) T. Karras, S. Laine, and T. Aila A style-based generator architecture for generative adversarial networks. In CVPR, Cited by: AI image generation.. Liu et al. (2026) R. Liu, M. Cui, Z. Qin, Z. Yan, R. Chen, Y. Han, Z. Li, J. Chen, Z. Chen, K. Lin, J. Shen, L. Weng, J. Dong, Y. Wang, and S. Wu MIRROR: manifold ideal reference reconstructor for generalizable AI-generated image detection. External Links: 2602.02222 Cited by: Table 4. Liu et al. (2020) Z. Liu, X. Qi, and P. H. S. Torr Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 8057–8066. Cited by: AI-generated image detection., Signal-level probes.. Madaan et al. (2023) A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y. Yang, S. Gupta, B. P. Majumder, K. Hermann, S. Welleck, A. Yazdanbakhsh, and P. Clark Self-refine: iterative refinement with self-feedback. In NeurIPS, Cited by: Self-improving agents and agentic forensics.. Ojha et al. (2023) U. Ojha, Y. Li, and Y. J. Lee Towards universal fake image detectors that generalize across generative models. In CVPR, Cited by: Introduction, AI-generated image detection.. OpenAI (2025) OpenAI Introducing 4o image generation. Note: https://openai.com/index/introducing-4o-image-generation/Accessed: 2026-06-30 Cited by: AI image generation.. OpenAI (2026) OpenAI Introducing chatgpt images 2.0. Note: https://openai.com/index/introducing-chatgpt-images-2-0/Accessed: 2026-06-30 Cited by: AI image generation.. Park and Owens (2024) J. Park and A. Owens Community forensics: using thousands of generators to train fake image detectors. External Links: 2411.04125 Cited by: Introduction, AI-generated image detection., Model-based experts., Table 2, Table 3, Table 4, Table 5. Peebles and Xie (2023) W. Peebles and S. Xie Scalable diffusion models with transformers. In ICCV, Cited by: AI image generation.. Ponomarenko et al. (2015) N. Ponomarenko, L. Jin, O. Ieremeiev, V. Lukin, K. Egiazarian, J. Astola, B. Vozel, K. Chehdi, M. Carli, F. Battisti, and C.-C. J. Kuo Image database tid2013: peculiarities, results and perspectives. Signal Processing: Image Communication. Cited by: Benchmarks.. Rombach et al. (2022) R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer High-resolution image synthesis with latent diffusion models. In CVPR, Cited by: AI image generation.. Sheikh et al. (2006) H. R. Sheikh, M. F. Sabir, and A. C. Bovik A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing. Cited by: Benchmarks.. Shinn et al. (2023) N. Shinn, F. Cassano, E. Berman, A. Gopinath, K. Narasimhan, and S. Yao Reflexion: language agents with verbal reinforcement learning. In NeurIPS, Cited by: Self-improving agents and agentic forensics.. Tan et al. (2024) C. Tan, Y. Zhao, S. Wei, G. Gu, P. Liu, and Y. Wei Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 28130–28139. Cited by: Table 3. Tan et al. (2025) H. Tan, J. Lan, Z. Tan, A. Liu, C. Song, S. Shi, H. Zhu, W. Wang, J. Wan, and Z. Lei Veritas: generalizable deepfake detection via pattern-aware reasoning. External Links: 2508.21048 Cited by: Introduction, Introduction, AI-generated image detection., Model-based experts., Benchmarks., Table 2, Table 3, Table 4, Table 5, Table 6. Wang et al. (2020) S. Wang, O. Wang, R. Zhang, A. Owens, and A. A. Efros CNN-generated images are surprisingly easy to spot… for now. In CVPR, Cited by: Introduction, AI-generated image detection., Signal-level probes.. Wang et al. (2023) Z. Wang, J. Bao, W. Zhou, W. Wang, H. Hu, H. Chen, and H. Li DIRE for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), p. 22445–22455. Cited by: AI-generated image detection.. Wen et al. (2025) S. Wen, J. Ye, P. Feng, H. Kang, Z. Wen, Y. Chen, J. Wu, W. Wu, C. He, and W. Li Spot the fake: large multimodal model-based synthetic image detection with artifact explanation. In Advances in Neural Information Processing Systems, Cited by: Table 4, Table 5. Wu et al. (2025a) C. Wu, J. Li, J. Zhou, J. Lin, K. Gao, K. Yan, S. Yin, S. Bai, X. Xu, Y. Chen, et al. Qwen-image technical report. External Links: 2508.02324 Cited by: AI image generation.. Wu et al. (2025b) C. Wu, P. Zheng, R. Yan, S. Xiao, X. Luo, Y. Wang, W. Li, X. Jiang, Y. Liu, J. Zhou, et al. OmniGen2: exploration to advanced multimodal generation. External Links: 2506.18871 Cited by: AI image generation.. Xu et al. (2025) Z. Xu, X. Zhang, R. Li, Z. Tang, Q. Huang, and J. Zhang FakeShield: explainable image forgery detection and localization via multi-modal large language models. In International Conference on Learning Representations, Cited by: Table 4. Yan et al. (2025a) S. Yan, O. Li, J. Cai, Y. Hao, X. Jiang, Y. Hu, and W. Xie A sanity check for ai-generated image detection. In International Conference on Learning Representations, External Links: Link Cited by: Introduction, Introduction, AI-generated image detection., Model-based experts., Benchmarks., Table 2, Table 3, Table 3, Table 4, Table 5. Yan et al. (2025b) Z. Yan, J. Wang, P. Jin, K. Zhang, C. Liu, S. Chen, T. Yao, S. Ding, B. Wu, and L. Yuan Orthogonal subspace decomposition for generalizable AI-generated image detection. In Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 267, p. 70268–70288. Cited by: Table 4. Yang et al. (2026) Y. Yang, Z. Gong, W. Huang, Q. Yang, Z. Zhou, Z. Huang, Y. Li, X. Gao, Q. Dai, B. Liu, K. Qiu, Y. Yang, D. Chen, X. Yang, and C. Luo SkillOpt: executive strategy for self-evolving agent skills. External Links: 2605.23904 Cited by: Self-improving agents and agentic forensics., Stage I: Tool-Use Skill Learning. Yao et al. (2023) S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao ReAct: synergizing reasoning and acting in language models. In ICLR, Cited by: Self-improving agents and agentic forensics.. Ye et al. (2025) J. Ye, B. Zhou, Z. Huang, J. Zhang, T. Bai, H. Kang, J. He, H. Lin, Z. Wang, T. Wu, Z. Wu, Y. Chen, D. Lin, C. He, and W. Li LOKI: a comprehensive synthetic data detection benchmark using large multimodal models. In International Conference on Learning Representations, External Links: Link Cited by: Introduction, Benchmarks.. Zhang et al. (2025) W. Zhang, C. Jiang, Z. Zhang, C. Si, F. Yu, and W. Peng IVY-fake: a unified explainable framework and benchmark for image and video aigc detection. arXiv preprint arXiv:2506.00979. Cited by: Table 3. Zhao et al. (2024) A. Zhao, D. Huang, Q. Xu, M. Lin, Y. Liu, and G. Huang ExpeL: llm agents are experiential learners. In AAAI, Cited by: Self-improving agents and agentic forensics., Stage I: Tool-Use Skill Learning. Zheng et al. (2025) B. Zheng, M. Y. Fatemi, X. Jin, Z. Z. Wang, A. Gandhi, Y. Song, Y. Gu, J. Srinivasa, G. Liu, G. Neubig, and Y. Su SkillWeaver: web agents can self-improve by discovering and honing skills. External Links: 2504.07079 Cited by: Self-improving agents and agentic forensics., Stage I: Tool-Use Skill Learning. Zhong et al. (2023) N. Zhong, Y. Xu, S. Li, Z. Qian, and X. Zhang PatchCraft: exploring texture patch for efficient ai-generated image detection. arXiv preprint arXiv:2311.12397. External Links: Link Cited by: Introduction, Benchmarks.. Zhong et al. (2025) N. Zhong, M. Zou, Y. Xu, Z. Qian, X. Zhang, B. Wu, and K. Ma Self-supervised ai-generated image detection: a camera metadata perspective. External Links: 2512.05651 Cited by: Introduction, AI-generated image detection., Model-based experts., Table 2, Table 3, Table 4, Table 5. Zhu et al. (2026) C. Zhu, M. Wang, J. Liu, C. Chang, and I. Echizen EvoGuard: an extensible agentic rl-based framework for practical and evolving ai-generated image detection. External Links: 2603.17343 Cited by: Table 4, Table 4.

Instant research discovery