acceptodds
Under review as a conference paper at ICLR 2027

EvoGuard: An Extensible Agentic RL-based Framework for Practical and Evolving AI-Generated Image Detection

Abstract

The rapid proliferation of AI-Generated Images (AIGIs) poses severe misinformation risks, making AIGI detection critical yet challenging. Traditional detection paradigms mainly rely on low-level features, whereas recent research increasingly focuses on leveraging the general understanding ability of Multimodal Large Language Models (MLLMs) to achieve better generalization, yet it still suffers from limited extensibility and expensive data annotations. Instead of building yet another detector, we recast AIGI detection as learned, reasoning-based evidence synthesis over a pool of heterogeneous off-the-shelf detectors, realized through EvoGuard, a novel agentic framework. A capability-aware selection mechanism profiles each detector and gathers complementary evidence per sample; a dynamic orchestration mechanism then reasons over heterogeneous outputs across multiple rounds, cross-validating conflicting or low-confidence signals before concluding. This design exploits the complementary strengths among heterogeneous detectors, transcending the limits of any single model. Furthermore, optimized by a GRPO-based Agentic Reinforcement Learning algorithm using only low-cost binary labels, it eliminates the reliance on fine-grained annotations. Extensive experiments demonstrate that this learned reasoning paradigm outperforms single-detector and static ensembling, achieving SOTA accuracy while mitigating the bias between positive and negative samples. More importantly, it allows the plug-and-play integration of new detectors to boost overall performance in a train-free manner, offering a highly practical, long-term solution to ever-evolving AIGI threats. Source code is included in the supplementary materials and will be publicly available upon acceptance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-Generated Image Detection, Agentic AI, Image Forensics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_XfTHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EvoGuard, an extensible agentic framework for AI-generated image detection that integrates heterogeneous detectors through a multimodal large language model. It introduces capability-aware tool selection and dynamic orchestration to adaptively invoke detectors, analyze their outputs, and synthesize evidence for final predictions. The agent is trained using GRPO with only image-level binary labels, avoiding the need for fine-grained reasoning annotations. Experiments on three in-the-wild benchmarks demonstrate competitive detection performance, improved balance between real and fake image accuracy.

Critical strengths

1. An interesting and practical framework for heterogeneous detector integration.

The paper presents an interesting agentic framework that reformulates AI-generated image detection as evidence synthesis over heterogeneous off-the-shelf detectors, rather than developing another standalone detector. The combination of capability-aware tool profiles and dynamic orchestration provides a flexible approach to integrating detectors with different architectures and output formats. This is a meaningful direction for addressing the limitations of individual detectors in evolving real-world scenarios.

2. Promising extensibility with relatively low training annotation requirements.

A notable advantage of EvoGuard is its ability to incorporate additional detectors without retraining the agent. The experiments on training-free tool extension support this practical benefit, which is particularly relevant given the rapid evolution of generative models. In addition, the GRPO-based training strategy relies on image-level binary labels rather than expensive fine-grained reasoning annotations, making the framework potentially easier to adapt and maintain as new detection methods become available.

Critical weaknesses

While EvoGuard demonstrates promising performance by integrating heterogeneous detectors, the current experiments do not sufficiently disentangle the benefits of detector complementarity from those of learned agentic decision-making. In Table 3, the select one tool variant achieves an average B-Acc of 92.00%, nearly identical to MIRROR alone (91.97%), raising questions about whether the agent genuinely learns sample-adaptive tool selection or predominantly relies on the strongest detector. More importantly, the one round variant, which allows more flexible tool selection, consistently underperforms select one tool across all three benchmarks (90.67% vs. 92.00% average B-Acc). Although the full framework improves performance to 92.70%, the paper does not adequately explain these observations or demonstrate how the learned orchestration strategy effectively selects and integrates complementary evidence. Consequently, the current evidence remains insufficient to fully substantiate the claimed advantages of intelligent tool selection and multi-round reasoning over simpler detector selection and integration strategies.

Questions for authors

1. Does the agent genuinely learn sample-adaptive tool selection?

The select one tool variant performs almost identically to MIRROR alone. Could the authors report the distribution of first-round tool selections across different image categories? In particular, it would be valuable to examine whether the agent frequently defaults to MIRROR or selects alternative detectors when they are more reliable for specific samples. Comparing the accuracy of the selected detectors against MIRROR on the same samples would provide more direct evidence of adaptive tool selection.

2. Why does flexible single-round tool selection underperform selecting a single tool?

An unexpected observation in Table 3 is that one round consistently performs worse than select one tool across all three benchmarks, despite allowing more flexible tool selection. Could the authors explain this degradation? Does invoking additional detectors introduce conflicting or misleading evidence, or does the limitation arise from the agent's ability to integrate multiple outputs within a single round? It would also be helpful to clarify how subsequent reasoning rounds in the full EvoGuard framework mitigate this issue.

3. Does dynamic orchestration adapt to sample difficulty and effectively resolve detector conflicts?

Figure 4 reports the overall distributions of interaction rounds and tool calls, but does not reveal how these behaviors relate to sample difficulty. Could the authors divide the evaluation samples according to the number of constituent detectors that correctly classify them (0–4) and report the corresponding termination rates, average tool calls, and final accuracy? For samples correctly classified by all four detectors, does the agent tend to terminate earlier with fewer tool invocations? Conversely, for challenging samples where three detectors are incorrect and only one is correct, can the agent identify and prioritize the reliable detector? Representative successful and unsuccessful decision trajectories would provide valuable insights into the learned orchestration behavior.

4. How does EvoGuard balance detection performance against inference cost?

EvoGuard improves the average B-Acc over MIRROR by 0.73 percentage points (92.70% vs. 91.97%), while requiring approximately 6.7 times the inference time on LOKI (17m35s vs. 2m37s). Although the improvements in class balance and extensibility are valuable, could the authors further clarify the practical benefits of this additional computational cost? A performance–cost analysis under different tool-call budgets, particularly against MIRROR and simpler sequential tool invocation strategies, would help demonstrate the efficiency of the proposed framework.

8 comments
AuthorsOpenReview ↗

Respond to Reviewer XfTH (1/6)

We thank the reviewer for the detailed review and for recognizing EvoGuard as an interesting and practical framework for integrating heterogeneous detectors, with promising train-free extensibility and low annotation requirements. We address each question below.

Question 1: First-round tool selection and comparison with MIRROR on the same images

Below we report the requested statistics for the 4-tool EvoGuard and the "select one tool" variant on all test images.

(1) Why the "select one tool" variant matches MIRROR. With a single call, the agent has to pick the detector it expects to be most reliable for the image, before seeing any tool output. In our tool pool, MIRROR is the most reliable detector in almost every category. Across the 18 (benchmark, content category) groups with at least 20 images, MIRROR is the most accurate detector in 13 and within 3 points of the best in 4 others. Accordingly, the variant selects MIRROR for 95.4% of images. For the remaining 4.6%, it selects Effort or FakeVLM, and on these same images the selected detector is about as accurate as MIRROR (91.5% vs. 93.6%). This explains why its B-Acc (0.9200) matches MIRROR's (0.9197). What a single call cannot do is correct the chosen detector when it is wrong, since no other evidence is available. This is what EvoGuard's multiple turns add, as shown in (3).

(2) First-round selection of EvoGuard by category. The table shows the first tool called by EvoGuard for each content category with at least 100 images, with the variant's MIRROR rate for reference.

Table R3-1.

Benchmark / categorynEvoGuard first call: MIRROR / FakeVLM / EffortSelect-one: MIRROR
LOKI / scenery45759 / 6 / 3585
LOKI / medical401100 / 0 / 0100
LOKI / animal39825 / 75 / 0100
LOKI / text screenshot32974 / 26 / 097
LOKI / object24897 / 1 / 2100
LOKI / portrait21555 / 44 / 1100
LOKI / satellite1401 / 0 / 9959
CF / object66391 / 3 / 698
CF / scenery55295 / 3 / 299
CF / portrait22875 / 24 / 1100
CF / human body17996 / 4 / 1100
CF / anime1034 / 96 / 035
Bfree / human body70395 / 5 / 0100
Bfree / portrait24180 / 13 / 6100
All5,06475.9 / 16.2 / 7.895.4

EvoGuard starts with MIRROR on 75.9% of images, compared with 95.4% for the variant, and its first call varies strongly with the category. For example, it starts with Effort on 99% of satellite images and with FakeVLM on 75% of LOKI animal images and 96% of CF anime images. This is not a fixed default either. In the 3-tool setting, where no single tool dominates, the first call is split between FakeVLM and Effort by category (e.g., FakeVLM for 88% of LOKI text screenshots and Effort for all LOKI satellite images). In the multi-turn setting, however, the first call is an entry point rather than the final choice: when EvoGuard starts with another tool, it still consults MIRROR afterwards on 98% of these images. The more informative comparison with MIRROR is therefore at the level of the final decision.

(3) Comparison with MIRROR on the same images.

  • When EvoGuard departs from MIRROR, it is usually right. Its final answer differs from MIRROR's on 192 images (3.8%). On these images, the evidence EvoGuard follows instead is correct 63.5% of the time (122 images), while MIRROR is correct 36.5% of the time (70 images).
  • Whether to override MIRROR depends on the sample. EvoGuard overrides MIRROR on 0.5% of images where MIRROR is confident (confidence ≥ 0.95). The rate rises to 3.8% at confidence 0.90–0.95 and to 11.2% below 0.90. In the last two groups, 65–67% of the overrides are correct. It uses a more reliable alternative where one exists. On LOKI text screenshots, the one category where another detector is clearly better (FakeVLM 84.8% vs. MIRROR 48.3%), EvoGuard calls FakeVLM on 98% of images and overrides MIRROR 59 times, 57 of them correctly, raising accuracy to 65.0% while the variant stays at 49.8% (T3 and T4 in our reply to Q3). Most other corrections recover fakes that MIRROR misses (net +98 on fakes vs. −46 on reals), hence its much smaller real/fake gap (0.006 vs. 0.054, Table 2).

(4) Summary. We agree that MIRROR is EvoGuard's main source of evidence in the 4-tool setting. This is warranted, since MIRROR is the most reliable detector in nearly every category, and it leaves limited headroom: MIRROR alone is correct on 90.3% of the images. Within this headroom, EvoGuard's gains come from deciding when to override MIRROR, which a single committed choice cannot do. As a result, EvoGuard outperforms the select-one variant on all three benchmarks, and achieves a much better balance between real images and fake images.

AuthorsOpenReview ↗

Respond to Reviewer XfTH (2/6)

Question 2: Why does "single round" underperform "select one tool", and how do later rounds help?

We thank the reviewer for this careful observation. To examine the two explanations raised by the reviewer, we grouped the test images by the agreement of the four detectors and compared the two variants and the full EvoGuard on the same images.

(1) Additional detectors are not harmful when they provide a majority. On images where the detectors split 3–1 (3 detectors hold the same judgement), the single-round variant performs on par with "select one tool" (B-Acc averaged over the three benchmarks: 0.868 vs. 0.869). It follows the majority on 88% of these images. The single-round variant is therefore able to integrate multiple outputs.

(2) The single-round variant falls behind on ties. On images where the detectors split 2–2, no majority exists, and the single-round variant is worse (0.6904 vs. 0.7708). Here it sides against MIRROR on 104 images, and these decisions are correct only 36% of the time. "Select one tool" almost always calls MIRROR alone, so it never sees the conflict. The difficulty is thus contradictory evidence, specifically evenly split evidence that a single reasoning step arbitrates poorly. The single-round constraint makes such conflicts frequent, because the agent must request all the tools it wants in that single round (3.5 calls per image on average, including the weakest detectors).

(3) How later rounds mitigate this. The full EvoGuard does not have to commit to its tools in advance. It sees the first outputs and then decides whether more evidence is needed (2.4 calls per image on average). It improves over the single-round variant in both conflict groups: from 0.8676 to 0.8858 on 3–1 splits and from 0.6904 to 0.7713 on ties. Its tie-breaking is also more reliable: when EvoGuard sides against MIRROR on a tie, it is correct 58% of the time, compared with 36% for the single-round variant, and it does so less often (74 vs. 104 images).

(4) What this ablation says about the design. EvoGuard is built around multi-round orchestration: the agent gathers evidence step by step and decides its next action based on what it has seen. The single-round variant keeps the same tools and reasoning but removes this loop. Without the loop, the extra detectors bring no gain on majority cases and hurt on ties, so the net result is below simply routing to the single most reliable detector (0.9067 vs. 0.9200). With the loop, the same detectors yield a gain over both variants (0.9270). Additional detectors therefore pay off only when the agent can act on their outputs, which is the motivation for our multi-round design.

AuthorsOpenReview ↗

Respond to Reviewer XfTH (3/6)

Quesiton 3: How does the agent's behavior relate to sample difficulty?

We grouped the test images of all three benchmarks by the number of detectors that classify them correctly (k = 0–4 correct detectors) and report the requested statistics for the 4-tool EvoGuard.

Table R3-2.

k (correct detectors)ImagesTerminates after round 1Avg. tool callsEvoGuard acc.MIRROR acc.Majority vote acc. (ties → MIRROR)
05757.9%2.230%0%0%
132650.3%2.3750.0%42.3%0%
292048.4%2.3083.8%82.6%82.6%
31,52646.4%2.3995.3%94.2%100%
42,23541.7%2.40100.0%100%100%
All5,06445.1%2.3891.3%90.3%89.3%

EvoGuard's advantage lies in the hard groups. For majority voting, we break 2–2 ties with MIRROR, the most accurate detector. Majority voting is always wrong when only one detector is right (k = 1), and on ties (k = 2) it gives the same answers as MIRROR (82.6%). EvoGuard reaches 50.0% and 83.8% in these two groups, and it beats the strongest single detector in every group where the detectors disagree. Overall, it reaches 91.3%, compared with 89.3% for majority voting and 90.3% for MIRROR.

(2) Does the agent stop earlier on easy images? The agent cannot know k (the number of "right" detectors), because k depends on the ground truth. What it can see is whether the outputs it has received so far agree, and its decision to stop follows this signal. When it has seen at least two outputs and still has unused tools, it calls more tools in 43.0% of cases if the outputs conflict, but in only 28.5% if they agree. Agreement among the detectors called so far is not the same as k, so the average number of calls is similar across groups (2.2–2.4), while accuracy differs greatly.

(3) Why do hard images end after the first round more often? In the table, 57.9% of the images at k = 0 end after the first round, compared with 41.7% at k = 4. This does not mean that the agent uses less evidence on hard images. On hard images, it more often calls several detectors at once in its first round (51% of the images at k = 0 and 43% at k = 1, vs. 36% at k = 4) and then usually answers. On easy images, it more often starts with one detector and adds others one at a time, which takes more rounds.

(4) When three detectors are wrong and one is right, can the agent find the right one? When three detectors agree and one disagrees, we call the one that disagrees the minority detector. Following it only helps if the agent can tell when it is right. We therefore compare two cases: the minority detector is right (k = 1) or wrong (k = 3).

Table R3-3.

Minority detectorMinority right (k = 1): images / % followedMinority wrong (k = 3): images / % followed
Any326 / 50.0%1,526 / 4.7%
MIRROR138 / 84.8%88 / 61.4%
FakeVLM136 / 33.8%477 / 3.4%
Effort or AIDE52 / 0%961 / 0.1%

“% followed” is the share of these images on which EvoGuard’s final answer agrees with the minority detector, whether or not it was called. Higher is better on the left, lower on the right.

The agent follows the minority detector about ten times more often when it is right than when it is wrong (50.0% vs. 4.7%). This holds at two levels. First, the agent trusts each detector according to how often that detector is right when it is in the minority. MIRROR is right in 61% of these cases (138 of 226), and the agent usually follows it. Effort and AIDE are right in only 5% of these cases (52 of 1,013), and the agent does not follow them, which is the better choice. Second, the agent also tells cases apart for the same detector. It follows FakeVLM in 33.8% of the images where FakeVLM is in the minority and right, but in only 3.4% of those where it is in the minority and wrong.

(5) Representative trajectories. We describe 6 LOKI trajectories: successes at k = 1 to 4, and failures at k = 0 to 1. In each case, the agent first reads the image tags and uses the tool profiles to choose which detectors to call. After each output, it decides whether to call more detectors or to answer.

Reviewer_noorOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes EvoGuard, an agentic framework for AI-generated image detection that reformulates the task as reasoning-based evidence synthesis over a set of heterogeneous off-the-shelf detectors, rather than training another standalone detector. The framework consists of two main components: Capability-Aware Selection, which characterizes each detector through tool profiles and selects suitable detectors according to the input image, and Dynamic Orchestration, which iteratively analyzes heterogeneous detector outputs, resolves conflicts, and decides whether additional tools should be invoked before producing the final prediction. The agent is trained with GRPO-based Agentic Reinforcement Learning using only image-level binary labels, with rewards for prediction correctness, output format, and sufficient analysis. Experimental results on LOKI, CommunityForensic-Eval, and Bfree-Test show that EvoGuard can outperform individual detectors and several ensemble baselines, while reducing the real/fake accuracy gap and supporting train-free integration of additional detectors.

Critical strengths

Rather than improving a single detector or applying a fixed fusion rule, the paper treats existing detectors as tools and learns how to dynamically select, invoke, and reason over them. This formulation is well motivated by the rapidly evolving nature of generative models and offers a potentially useful direction for maintaining an extensible detection system. In particular, the multi-round decision process is more flexible than conventional routing or static ensembling.

EvoGuard is designed to reason jointly over detectors with substantially different architectures and output formats, including numerical confidence scores and textual forensic explanations. The experiments show that the complete four-tool setting achieves an average B-Acc of 0.9270 and F1 of 0.9337, outperforming its constituent detectors while also reducing the average real/fake accuracy gap to 0.006. This supports the claim that the agent can exploit complementary detector behavior instead of simply inheriting the bias of a single model.

Critical weaknesses

The central claim of the paper is that EvoGuard can integrate future detectors in a train-free manner. However, the reported experiment simulates this by withholding existing detectors during training and adding them back at inference time, while the newly added tool is still drawn from the same detector pool considered by the authors. Moreover, a manually specified Overall Profile is still provided for the new detector. Therefore, it remains unclear whether the same behavior would hold for genuinely novel future detectors with substantially different failure modes, output schemas, or domains.

The Tool Profiles are central to Capability-Aware Selection, yet the paper states that the profiles are first generated from training-set statistics by an LLM and then manually refined. Since the ablation only removes the profile entirely, it does not disentangle the contribution of automatically computed statistics from the contribution of human refinement. This raises concerns about reproducibility and the actual deployment cost of adding new tools. In particular, if substantial human expertise is needed to prepare Conflict Hints and detector strengths/weaknesses, the claimed practicality and extensibility become less clear.

Questions for authors

Q1. How important is the manual refinement step in Tool Profile construction? Could the authors report an ablation comparing: <1>fully automatically generated profiles, <2>automatically generated + manually refined profiles, <3>only Overall Profile, <4>the complete profile with Strengths / Weaknesses / Conflict Hints? This would help clarify the actual human cost required to add a new detector.

Q2. How does EvoGuard behave when integrating a truly unseen detector with a substantially different output format or failure distribution? The current extensibility experiment uses held-out detectors from the existing pool. Could the authors clarify whether the agent has been tested with a detector whose architecture, output schema, or forensic principle was entirely absent during training?

Q3. Why are AlignGemini and X2-DFD not included in the empirical comparison? Since Table 1 identifies them as methods that can combine heterogeneous detectors and potentially support train-free extension, they appear highly relevant to EvoGuard's central claim. If official implementations are unavailable, a discussion explaining why they cannot be evaluated would be helpful.

Q4. Can the authors provide an accuracy-versus-computation comparison? For example, it would be useful to report performance under: <1>one tool call, <2>at most two tool calls, <3>one round, <4>two rounds, <5>full dynamic orchestration, together with average tool calls, FLOPs or latency per image. This would better isolate the benefit of reasoning from the benefit of additional test-time compute.

5 comments
AuthorsOpenReview ↗

Respond to Reviewer noor (1/5)

We thank the reviewer for the thoughtful review and for recognizing the motivation and flexibility of EvoGuard. We address each question below.


Question 1: "How important is manual refinement of the tool profiles?"

We thank the reviewer for this question, since the cost of adding a detector is central to our extensibility claim. In short, manual refinement is light and not required. Fully automatic profiles reach comparable performance, while our refined full profiles give the best F1 and class balance. A held-out detector can be added without retraining using only its one-line overall profile.

11 How profiles are built and what the manual pass changed. Profile generation is automatic up to the last step. The backbone MLLM tags each training image, we compute each tool's accuracy and false positive/negative rates per tag, and an LLM drafts the profile from these statistics. The manual pass was light and was done by one author in a short time:

  • it expanded the abbreviations FPR/FNR;

  • it listed MIRROR's weak tags explicitly;

  • it noted that AIDE's hints may not hold for medical and remote-sensing images, which are rare in the training set;

  • it added "yet powerful" to FakeVLM's overall profile;

  • it added two general notes: the definition of false positives/negatives, and that profiles are derived from the training set.

These edits mainly clarify the wording and use no information beyond the training-set statistics. The final profiles are given in Appendix A.10.

22 Four-way ablation. Besides the paper model automatic+manualrefinement,fullprofileautomatic + manual refinement, full profile, we trained the 4-tool EvoGuard with the other three variants. "Full" means overall profile + strengths + weaknesses + conflict hints. The R/F gap is the R/F Acc Gap of Table 2.

Table R2-1

ProfileB-AccF1R/F-Acc gap
Fully automatic, full0.92200.92750.023
Fully automatic, overall only0.91740.92770.029
Automatic + manual, overall only0.92580.92940.038
Automatic + manual, full (paper)0.92700.93370.006

B-Acc varies little across the four variants 09˙17–09˙270\.917–0\.927. Our refined full profile gives the best F1 and by far the most balanced decisions, which is the bias mitigation reported in Sec. 4.2.

  • With overall-only profiles, the R/F gap rises to 0.029–0.038, mainly because more fakes are missed on LOKI F-Acc 0\.791–0\.809 vs\. 0\.861. The strengths, weaknesses and conflict hints are what keep EvoGuard's decisions balanced.

  • To compare profiles on the very same model, we also replaced the paper model's profiles with the fully automatic overall-only ones at test time, without retraining. B-Acc is unchanged 0\.9278 vs\. 0\.9270, but the R/F gap rises from 0.006 to 0.059.

  • Removing profiles entirely gives 0.8671 B-Acc (Table 3).

Hence our refined full profile gives the best overall result. Fully automatic profiles already reach comparable B-Acc, so the framework does not depend on the manual pass.

33 Cost of adding a new detector. Following the protocol of Sec. 4.3, the held-out tools receive only their one-line overall profiles, and the agent is not retrained. As shown in Fig. 3 (see also our reply to Q2), this already recovers most of the gain of training with all four tools.

In practice, adding a detector therefore requires no retraining. A one-line description of the tool already recovers most of the gain Fig\. 3, and a full profile can be generated automatically by running the detector once on the training set; a brief manual check of this profile could further improves class balance, as shown in (2).

AuthorsOpenReview ↗

Respond to Reviewer noor (2/5)

Question 2: "Integrating a truly unseen detector?"

We thank the reviewer for this question. The held-out detectors in our extensibility experiments Sec\. 4\.3, Fig\. 3 are entirely absent during training, so to the agent they are as new as a detector released after training. They also differ substantially from the training tools in architecture, forensic principle and failure distribution. Adding them at test time without retraining recovers 90–96% of the gain of training with all four tools.

11 The held-out detectors are new to the agent. Although the four detectors come from our pool, the agent never encounters a held-out detector during training:

  • it is absent from the tool list, so the agent can never call it;

  • it has no profile;

  • none of its outputs appear in any training trajectory.

At test time we only add it to the tool list and describe it in the system prompt, without any parameter update.

The four detectors also differ substantially from each other. R-Acc / F-Acc are measured for each detector alone, averaged over the three benchmarks.

Table R2-2

EffortAIDEFakeVLMMIRROR
Descriptionclassifier on CLIP featuresclassifier on DCT frequency + CLIP featuresfine-tuned MLLM that explains artifactsmanifold consistency on DINO features
R-Acc / F-Acc alone0.822 / 0.6950.874 / 0.4460.576 / 0.9300.947 / 0.892

The four detectors span three architectures CLIP−basedclassifiers,afine−tunedMLLMandaDINO−baseddetectorCLIP-based classifiers, a fine-tuned MLLM and a DINO-based detector and different forensic principles, and their failures differ sharply: Effort and AIDE miss many fakes, FakeVLM flags about 42% of real images as fake, and MIRROR rarely raises false alarms.

22 The extensibility experiments in the paper Fig\. 3. Each of the three settings in Fig. 3 trains the agent with a subset of the tools and adds the held-out tools at test time without retraining. In every setting, adding the unseen detectors without retraining recovers 90–96% of the gain that training with all four tools brings over using the training tools alone Fig\. 3. The agent also uses the unseen detector much as an agent trained with it does: in cc, it overturns about 30% of MIRROR's misses on fake images, compared with 31.2% for the agent trained with MIRROR.

Like API documentation, any tool needs at least a name and a one-line description of its purpose to be callable. The tag-level profile helps further, and it comes from training-set statistics rather than from the authors' knowledge.

AuthorsOpenReview ↗

Respond to Reviewer noor (3/5)

Question 3: Comparison with AlignGemini and X2-DFD

We agree that both methods are closely related to EvoGuard, since both combine heterogeneous detectors. The core design of AlignGemini is already evaluated in Table 3, and below we give more complete results. The official X2-DFD checkpoint was released in June 2026, after our main experiments were completed; we add the comparison here.

11 AlignGemini: its fusion design is evaluated in Table 3. AlignGemini fuses an MLLM branch Qwen25˙−VL−7BQwen2\.5-VL-7B and a DINOv2-based branch with an OR rule: an image is labeled fake if either branch flags it. FakeVLM ⊕ MIRROR in Table 3 (⊕ denotes OR) applies this rule to off-the-shelf detectors of the same two types. Since we found no official implementation, this evaluates the fusion rule rather than AlignGemini’s own branches. Below we extend it to all OR combinations of FakeVLM with a non-MLLM detector:

Table R2-3

Method B−AccB-AccLOKICommunityForensicBfree-TestAverage
FakeVLM ⊕ Effort0.75530.74530.67220.7243
FakeVLM ⊕ AIDE0.78650.74630.67560.7361
FakeVLM ⊕ MIRROR (Table 3)0.80030.76350.70470.7562
MIRROR alone0.85230.93320.97360.9197
EvoGuard (4 tools)0.86380.93800.97920.9270

Two observations follow. First, OR fusion drags down even the strongest detector: FakeVLM ⊕ MIRROR is 16 points below MIRROR alone, because the OR rule inherits every false positive of both detectors realaccuracydropsto04˙2–06˙5real accuracy drops to 0\.42–0\.65. Second, although OR fusion can add detectors without training, each added detector can only keep or lower real accuracy, so its extensibility comes at the cost of more false positives. EvoGuard instead reasons over the detector outputs and outperforms every OR combination on all three benchmarks.

22 X2-DFD: comparison added. We evaluated the official X2-DFD checkpoint on our three benchmarks with its own two expert detectors. X2-DFD passes the scores of non-MLLM detectors to an MLLM in a single pass.

Table R2-4

MethodLOKICommunityForensicBfree-TestAverage
X2-DFD, B-Acc threshold05˙threshold 0\.50.5000.4990.5080.502
X2-DFD, B-Acc thresholdtunedby5−foldCVthreshold tuned by 5-fold CV0.5280.6140.7250.622
EvoGuard (3 tools), B-Acc0.80350.81250.80440.8068
EvoGuard (4 tools), B-Acc0.86380.93800.97920.9270

With the default threshold of 0.5, X2-DFD labels almost every image as fake. Even with a decision threshold tuned by 5-fold cross-validation, which favors X2-DFD, its average B-Acc is 0.622, far below EvoGuard with either 3 tools 08˙0680\.8068 or 4 tools 09˙2700\.9270. Domain mismatch is unlikely to be the only cause: although X2-DFD was originally designed for face deepfake detection, even on Bfree-Test, where nearly all images depict people, it reaches only 0.725, compared with 0.8044 and 0.9792 for EvoGuard.

To compare the two paradigms under identical conditions, we also refer to the "one round" variant in Table 3. It uses the same tools and backbone as EvoGuard but follows X2-DFD's single-pass design: the agent invokes tools once and then answers. It reaches 0.9067 average B-Acc, below the multi-round EvoGuard 09˙2700\.9270, which shows the benefit of iterative orchestration over one-shot aggregation of detector outputs.

Reviewer_aBRdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes EvoGuard, an agentic framework for AIGI detection. By wrapping off-the-shelf SOTA AIGI detectors as executable tools and carefully crafting the tool profiles, EvoGuard enables the MLLM-based agent to invoke appropriate tools based on the style and subject of the image being detected. Furthermore, EvoGuard employs GRPO to allow the agent to explore tool-calling strategies. Experimental results demonstrate its effectiveness.

Critical strengths

AIGI Detectors as Tools. Encapsulating AIGI detectors as tools effectively addresses the routing and voting difficulty caused by inconsistent output spaces among detectors.

Critical weaknesses
  1. Inconspicuous experimental results. As shown in Table 2, when only Effort + FakeVLM + AIDE are used as tools, EvoGuard underperforms compared to the corresponding best baselines on certain test sets. For instance, it performs worse than FakeVLM on LOKI and falls short of Effort on CommunityForensic-Eval. When Effort + FakeVLM + AIDE + MIRROR are all employed, EvoGuard's overall performance remains roughly on par with that of MIRROR.
  2. Lack of explanation for the hyperparameters in the analysis reward (e.g., the 10-character).
Questions for authors

Regarding the issue of performance being close to that of MIRROR, could you provide and analyze the invocation rates of EvoGuard for each tool when Effort + FakeVLM + AIDE + MIRROR are used as tools?

5 comments
AuthorsOpenReview ↗

Respond to Reviewer aBRd (1/3)

We thank the reviewer for the careful review and for recognizing that the EvoGuard paradigm effectively addresses the routing and voting difficulty caused by the inconsistent output spaces of different detectors. We address each point below.


Weakness 1: "The experimental results are not significant."

We agree that, with three tools, EvoGuard is not the best method on every benchmark. However, per-benchmark B-Acc alone gives an incomplete picture, and it does not reflect three advantages that matter in practice.

(1) EvoGuard beats an oracle that picks the best tool per benchmark. The two cases the reviewer notes amount to choosing a different tool for each benchmark in hindsight, a choice that requires test labels and cannot be made at deployment. Even this oracle selection reaches only 0.7835 average B-Acc, which is still 2.3 points below EvoGuard 08˙0680\.8068. Each of the two deficits is about 1.3 points. On Bfree-Test, by contrast, EvoGuard leads Effort by 9.6 points and FakeVLM by 11.6 points. Every single tool fails badly on at least one benchmark, while EvoGuard stays above 0.80 on all three:

Table R1-1

MethodAvg B-AccWorst-benchmark B-AccLowest R-/F-Acc on any benchmarkAvg R/F gap
Effort0.75840.70830.4630.200
FakeVLM0.75330.68890.4170.354
AIDE0.66010.51680.0870.428
Oracle tool per benchmark0.78350.70830.4630.228
EvoGuard (3 tools)0.80680.80350.6800.144

(2) EvoGuard mitigates one-sided failures. B-Acc is the average of real and fake accuracy, so it can mask severe class bias. On Bfree-Test, FakeVLM labels 58% of real images as fake R−Acc=04˙17R-Acc=0\.417, and Effort misses 54% of fake images F−Acc=04˙63F-Acc=0\.463. (see in Table. 2) EvoGuard reaches R-Acc=0.832 and F-Acc=0.777. EvoGuard reduces the average R/F gap of its tools by 28–66%. With 4 tools, this gap drops to 0.006, compared with 0.054 for MIRROR.

33 EvoGuard improves on a near-saturated MIRROR and absorbs it without retraining. MIRROR is already near saturation on CommunityForensic and Bfree-Test 09˙3–09˙7B−Acc0\.93–0\.97 B-Acc, yet EvoGuard improves on it on all three benchmarks in both B-Acc and F1. On LOKI, MIRROR's weakest benchmark, EvoGuard improves B-Acc by 1.2 points and F1 by 1.8 points, raising F-Acc from 0.822 to 0.861. More importantly, EvoGuard is not another detector competing with MIRROR. It is a framework built on top of detectors, and its ceiling rises as the toolbox improves:

  • It absorbs a stronger tool without retraining. In Fig. 3cc, an agent trained only with the three earlier tools reaches 0.8068. Adding MIRROR at test time raises this to 0.9209. Without any retraining, it already reaches the level of MIRROR alone 09˙1970\.9197 and recovers about 95% of the gain from full 4-tool training 09˙2700\.9270.

  • Weaker tools do not drag it down. With 4 tools, EvoGuard exceeds MIRROR on all three benchmarks in both B-Acc and F1. In contrast, every fusion baseline in Table 3 falls below MIRROR in B-Acc on every benchmark: gating network, weighted voting, score averaging, the calibrated variants, and logical OR. Even our train-free extension 09˙2090\.9209 outperforms all of them best:09˙108best: 0\.9108, although they use all four tools.

This is the core motivation of EvoGuard. No single detector stays the best for long, and with a fixed detector or a fixed fusion rule, every new detector requires re-selecting the deployed model or re-fitting the rule, while naive fusion can even fall below the strongest member (Table 3). EvoGuard incorporates a new detector by adding its description without retraining, so its performance keeps pace with the detector pool.

AuthorsOpenReview ↗

Respond to Reviewer aBRd (2/3)

Weakness 2: "The hyperparameters in the analysis reward eg˙,˙10characterse\.g\., 10 characters lack justification."

We thank the reviewer for this question. This threshold is not a tuned hyperparameter used to control reasoning length. Instead, it serves as a criterion to prevent the agent from outputting a prediction without analyzing.

11 What the thresholds check. After tags and whitespace are removed, the analysis reward checks two parts of each trajectory. The text before the first tool call must exceed 10 characters, and the text between the last tool response and the final answer must exceed 20 characters. Each failed check incurs a penalty of −0.2. These floors only rule out empty or single-word outputs. They do not reward longer reasoning, and the form of the reasoning is left unconstrained.

22 What the reward prevents. Without it, the agent never writes any analysis before its first tool call 00% of test trajectories, vs\. 100% for EvoGuard, and B-Acc drops on all three benchmarks −13˙to−45˙points;average09˙270→09˙027,Table3−1\.3 to −4\.5 points; average 0\.9270 → 0\.9027, Table 3, even though the agent then calls more tools 3\.46 vs\. 2\.38 per image.

(3) The exact threshold values are not sensitive. We retrained EvoGuard with 5× larger thresholds 50/10050/100 under the same training setting and compared the analysis length across the 5,064 test trajectories. The table reports median lengths in characters. Total analysis counts all text the agent writes, excluding tool calls and tool responses.

Table R1-2

Analysis rewardAnalysis before the first tool callTotal analysisAvg B-Acc
None (Table 3)01,1020.9027
Thresholds 10 / 20 (ours)8691,8280.9270
Thresholds 50 / 1008761,8260.9231
  • Raising the thresholds five-fold leaves the analysis length largely unchanged 869 vs\. 876 characters before the first tool call; 1,828 vs\. 1,826 in total.

  • Without the reward, the agent calls tools before examining the image and reasons only after seeing their outputs. Its total analysis is 40% shorter.

  • B-Acc changes only slightly 0\.9231 vs\. 0\.9270, −0\.4 points, much less than removing the reward −24˙points−2\.4 points, and remains above MIRROR alone 09˙1970\.9197 and every other variant in Table 3.

So although our thresholds are small, they effectively elicit analysis of adequate length. They only need to rule out skipping the analysis; once the analysis is written, its length does not depend on the threshold value.

AuthorsOpenReview ↗

Respond to Reviewer aBRd (3/3)

Question 1: Invocation rate of each tool

We thank the reviewer for this suggestion. Below we report the invocation rate of each tool for the 4-tool EvoGuard on all three benchmarks, and analyze how the agent uses them. On average, EvoGuard 4−tools4-tools makes 2.38 calls per image.

Tool avgB−Accaloneavg B-Acc aloneLOKICommunityForensicBfree-TestOverall
MIRROR 09˙1970\.919799.9%99.8%99.3%99.8%
FakeVLM 07˙5330\.753378.4%81.6%84.7%80.8%
Effort 07˙5840\.758443.6%45.6%35.5%42.8%
AIDE 06˙6010\.660117.7%11.0%12.2%14.2%

11 The invocation pattern follows reliability and complementarity. EvoGuard uses MIRROR, the strongest tool, as its anchor and calls AIDE, the weakest, least often. Notably, FakeVLM is called almost twice as often as Effort despite similar standalone B-Acc. A likely reason is that FakeVLM covers MIRROR's main weakness. On all three benchmarks, MIRROR misses more fakes than it misclassifies reals e\.g\., F-Acc 0\.822 vs\. R-Acc 0\.882 on LOKI, whereas FakeVLM has high fake accuracy 08˙89–09˙610\.889–0\.961. The agent has therefore learned to pair the strongest tool with the one that covers its blind spot, rather than calling tools uniformly.

(2) EvoGuard does not simply copy MIRROR. It consults at least one other tool on about 89% of images. It corrects 122 of MIRROR’s 493 errors 247˙24\.7% while overturning only 70 of its 4,571 correct predictions 15˙1\.5%, mostly by recovering fakes that MIRROR missed. This is why EvoGuard is more balanced than MIRROR: on LOKI, F-Acc rises from 0.822 to 0.861, and the average R/F gap on 3 test sets drops from 0.054 to 0.006 (Table 2).

33 The headroom is limited when a strong tool already exists. MIRROR alone is already correct on 90.3% of these images, so the room for improvement is inherently small. Within this room, EvoGuard follows it where it is reliable and overrides it where complementary evidence disagrees. As discussed in our response to W1, anchoring on the strongest available tool without being dragged down by weaker ones is exactly what the framework is designed to do, and it lets EvoGuard benefit directly from future, stronger detectors.


Revisions. In the final version, we will add the per-benchmark analysis TableR1−1Table R1-1 to Appendix A.8, the implementation details of the analysis reward and the threshold-sensitivity results TableR1−2Table R1-2 to Appendix A.1, and the tool invocation rates to Appendix A.5. We hope these responses address the reviewer’s concerns and are glad to answer any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Summary for the AC and Reviewers

Dear AC and Reviewers,

We thank all reviewers for the constructive discussion. We summarize the main results below.

Recognized contributions

  • Reformulating AIGI detection as reasoning-based evidence synthesis. EvoGuard treats existing detectors as tools and learns to dynamically select, invoke and reason over them (aBRd, noor, XfTH). Its multi-round decision process is more flexible than conventional routing or static ensembling (noor).
  • Strong performance by integrating heterogeneous detectors. EvoGuard resolves the routing and voting difficulties caused by the detectors' inconsistent output spaces (aBRd). By exploiting their complementary behavior, it achieves the best performance and makes the real/fake accuracy gap 9–70 times smaller than that of the detectors it uses (0.006 vs. 0.054–0.428), giving much more balanced predictions (noor, XfTH).
  • Extensibility and practical potential. Experiments show that EvoGuard can integrate additional detectors without retraining the agent and still improve performance. Its training is also cheap, requiring only image-level binary labels. This lets it keep pace with the rapid evolution of generative models and offers a promising long-term direction for maintaining an extensible detection system (noor, XfTH).

Reviewer feedback

  • Reviewer aBRd acknowledged the additional experimental results and our analysis of the reward hyperparameters, and noted Reviewer noor's concern about the impact of manually refined tool profiles. Our new ablation shows that manual refinement is light and not required: fully automatic profiles already reach a B-Acc of 0.9220, close to the 0.9270 of the refined profiles. Refinement mainly improves class balance.
  • Reviewer noor has not yet responded. We have tried to address the raised concerns with additional empirical evidence: an ablation of manual profile refinement, comparisons with AlignGemini and X2-DFD, and an analysis of accuracy versus test-time compute. We also explained that the detectors added in our extensibility experiment are completely unseen during training and differ substantially from the training tools.
  • Reviewer XfTH found our explanations convincing. The additional analyses of tool selection, multi-round orchestration and sample difficulty "address my main concerns and provide stronger evidence for the effectiveness of the proposed agent framework", and the performance–cost analysis is "valuable in demonstrating its practical potential". XfTH maintains the positive assessment.

Main additions

  1. The gain comes from reasoning, not from extra tool calls or the strongest detector (aBRd W1, noor Q4). With exactly the same tool calls on every image, EvoGuard beats the best fixed fusion rule by 1.5 points (0.9270 vs. 0.9116). Without MIRROR, it even beats an oracle that picks the best detector for each benchmark using test labels (0.8068 vs. 0.7835).
  2. Sample-adaptive tool selection and conflict resolution (XfTH Q1–Q3). Our analyses of tool selection, sample difficulty and detector conflicts show that the agent adapts to each image. When it overrides MIRROR, it is right 63.5% of the time and MIRROR only 36.5%. It follows a lone dissenting detector ten times more often when that detector is right than when it is wrong. These analyses also explain Table 3: a single round handles evenly split evidence poorly, and multi-round orchestration raises B-Acc on these cases from 0.6904 to 0.7713.
  3. Integration of truly unseen detectors (noor Q2). Detectors held out from training, absent from the tool list, the profiles and all training trajectories, can be added without retraining. They differ substantially from the training tools, and they recover 90–96% of the gain of retraining with all tools.
  4. Clear margins over related methods (noor Q3). On our detectors, AlignGemini's OR fusion reaches at most 0.7562, 17 points below EvoGuard (0.9270). X2-DFD, evaluated with its released checkpoint and its own expert detectors, reaches only 0.622. With the same tools and backbone as EvoGuard, X2-DFD's single-pass design (the "one round" variant of Table 3) still falls below multi-round EvoGuard (0.9067 vs. 0.9270).
  5. Lower cost without losing accuracy (noor Q4, XfTH Q4). Budget-aware training makes EvoGuard 1.6× faster with the same B-Acc (0.9271 vs. 0.9270), and adding a MIRROR-confidence gate makes it 2.2× faster (B-Acc 0.9275). At similar cost, gated EvoGuard is more accurate than a simple sequential rule.
  6. Justified analysis reward (aBRd W2). The 10- and 20-character thresholds are minimal floors that only rule out skipping the analysis. Without this reward, the agent writes no analysis before its first tool call, and B-Acc drops by 2.4 points despite more tool calls.

We would be grateful if the final evaluation could take into account these contributions, the reviewers' feedback, and the additional evidence summarized above. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.