acceptodds
Under review as a conference paper at ICLR 2027

Unlocking Native Forensic Representation in Multimodal Large Language Models

Abstract

Multimodal Large Language Models (MLLMs) have emerged as a promising paradigm for AI-generated image (AIGI) detection by leveraging high-level semantic understanding. However, reliable detection also depends on subtle low-level forensic evidence, motivating growing efforts to equip MLLMs with stronger forensic perception. This work demonstrates that strong forensic discrimination can be achieved using only the native vision representations of pretrained MLLMs. We introduce **Native Forensic Connection (NFC)**, a remarkably simple framework that directly connects native forensic representations to the language model. NFC is inspired by our observation that low-level forensic information is already highly discriminative in intermediate vision representations, but becomes substantially less accessible along the standard vision-language pathway. With NFC, MLLMs can access native low-level forensic information without adapting the pretrained vision encoder. To better evaluate low-level forensic perception, we further construct **SemClean-Bench** through a reverse filtering strategy that removes AIGIs with salient high-level semantic anomalies, shifting the evaluation toward low-level artifacts. Extensive experiments across diverse AIGI benchmarks and SemClean-Bench demonstrate strong and generalizable performance. These results show that strong forensic perception can be achieved by effectively accessing native visual representations without relearning, highlighting representation access as a simple and complementary direction for forensic MLLMs.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
AIGI Detection, AIGC Forensic, MLLM

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_osLQOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the issue of MLLM’s insufficient utilization of low-level forensic information in AIGC detection. The authors propose Native Forensic Connection (NFC), which connects forensically sensitive intermediate visual representations to specific layers of the language model via a lightweight alignment module. They also constructed the SemClean-Bench dataset to evaluate the model’s ability to recognize generated images in the absence of obvious semantic shortcuts. Experimental results show that NFC improves accuracy on SemClean-Bench, and when using a training setup closer to that of Deep-VRM, NFC-LoRA achieves 98.0% accuracy. The paper proposes a relatively concise method for integrating forensic information and investigates the low-level forensic capabilities of MLLMs from the perspectives of representation analysis and benchmark construction.

Critical strengths
  1. The research perspective is novel, proposing an approach to enhance forensic capabilities by accessing native intermediate visual representations.

  2. The evaluation design and ablation analysis are relatively comprehensive. SemClean-Bench constructs a targeted diagnostic benchmark to address potential shortcuts involving obvious semantic anomalies in existing generative image detection evaluations. The paper also evaluates detection performance on multiple conventional benchmarks and analyzes the effects of NFC connectivity, freezing the visual encoder, training order, fusion coefficients, and changes to the backbone model.

Critical weaknesses
  1. The filtering strategy used in SemClean-Bench is not yet sufficient to demonstrate that its evaluation results primarily reflect low-level forensic capabilities. The authors need to more clearly define the scope of conclusions that this benchmark can support.

  2. The experimental results cannot yet completely rule out the influence of data sources or fully demonstrate the generalization ability of NFC. The paper achieved a high average accuracy on SemClean-Bench, but there are significant differences in results among different generators.

Questions for authors
  1. Does SemClean-Bench truly measure low-level forensic capabilities?

  2. How can independence between the training data and the test benchmark be ensured? This primarily concerns the issue of information leakage.

  3. How should the performance differences of NFC across different generators be explained?

  4. Are the benefits of the two-stage training strategy statistically robust?

1 comment

Dear Reviewer osLQ,

Thank you for the constructive review. We appreciate your recognition of NFC’s use of native intermediate visual representations and the breadth of our evaluation.

[W1/Q1] Evaluation scope of SemClean-Bench

A1. Thank you for raising this point. SemClean-Bench evaluates detection performance when obvious semantic shortcuts are less available, offering a complementary perspective on forensic perception. Besides, we conduct an extra visual comparison of the image and attention maps between the excluded and retained images in SemClean-Bench. As shown in Fig.5, one can see that the retained part has better visual quality and more diffuse attention patterns. On the other hand, the excluded part has conspicuous semantic anomalies that attract concentrated model attention, suggesting more dependence on obvious semantic shortcuts. We also revise the manuscript to clarify the benchmark’s evaluation scope.


[W2/Q2/Q3] Data independence and Cross-generator generalization

A2. (a) Training–test independence. We strictly follow Deep-VRM's training and evaluation settings. Notably, the four recent generators included in our benchmark are released after the training dataset is constructed and are therefore unseen during training.

(b) Cross-generator generalization. For HiDream-O1, we observe a substantial drop in fake recall from 99.8% to 63.2% after applying random post-processing to the same images. This degradation may be attributed to the disruption of subtle low-level forensic cues by post-processing operations. Improving robustness against such transformations remains an important direction for future work.


[Q4] Robustness of Two-Stage Training

A3. Thank you for raising this question. We compare Stage 1 only, Stage 2 only, and both training orders, reporting the average and standard deviation over three runs with different random seeds. Our two-stage configuration achieves the best average performance across AIGI-Now, WildRF, and Chameleon.

Stage 1 onlyStage 2 onlyStage 2 → Stage 1Stage 1 → Stage 2
Avg. ± SD (%)74.4 ± 2.394.7 ± 0.373.1 ± 1.997.1 ± 0.2
Reviewer_37fjOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper reveals that the vision feature from shallow layer of VLM can be used for distinguishing AIGI, while the ability would be undermined during the LLM process. It introduces Native Forensic Connection (NFC) to inject the shallow visual features into selected LLM layers to recover the native low-level vision information for better classification. It also proposes SemClean-Bench to remain the AIGI with only low-level evidence. Experiments shows the effectiveness of NFC.

Critical strengths
  1. The paper is easy to understand and follow.
Critical weaknesses
  1. Limited contribution. The kernel claim that shallow layer of vision encoder is better for AIGI detection is a common realization in community. The ForensicsTok [1] and so many papers have revealed it. The paper don't provide extra information to make up for the consensus.

  2. Novelty. The operation that injecting vision features into selected layer in LLM is same like the deepstack in Qwen3, which has already been explored long ago.

  3. The SemClean-Bench comes from the existing datasets with a simple rule to filter the images with semantic abnormalities. It's difficult to approve that the contribution is enough.

[1] ForensicsTok: Forensics-Guided Tokenized Modeling for Image Tampering Localization

Questions for authors
  1. It's NFC needed to be customized for different VLM backbone? It's there any other results based on the different VLMs? It's interesting to find out if there is any common rule for different VLMs to distinguish which layer is the best for AIGC detection.

  2. The results in tables are lower than the reported in original paper. I wonder the difference while evaluate the other existing methods.

  3. Do you have the results for real image? The accuracy of real image is also a significant signal, which determines if the model could be used in reality.

3 comments

Dear Reviewer 37fj,

Thank you for your constructive review and for noting the clarity of the presentation. We address your concerns below.

[W1] Core Contribution.

A1. ForensicsTok focuses on forgery localization through autoregressive mask generation and enhances forensic perception by integrating external forensic experts. In contrast, our key finding is that the MLLM's own frozen vision encoder already contains effective forensic cues for AIGI detection, without requiring an additional forensic ViT. Based on this observation, NFC directly extracts and injects native forensic representations into the LLM, without introducing external experts or adapting the vision encoder. We will clarify these distinctions in the Related Work section.


[W2] Difference from DeepStack.

A2. Thank you for the suggestion. We investigate other common feature injection methods, e.g., DeepStack (Bai et al., 2025), developed for general visual understanding, and compare it with NFC using the same settings. NFC improves the average accuracy from 89.6 to 92.7 across three benchmarks, showing the benefit of intermediate feature extraction and alignment through our proposed NFC. Besides, our NFC is motivated by the strong forensic discrimination of the native visual representations in pretrained MLLMs. In contrast, Deepstack utilizes multiple connections to jointly optimize the ViT and LLM for better visual understanding.

SettingAIGI-NowWildRFChameleonAvg.
DeepStack88.090.690.389.6
NFC91.197.190.092.7

[W3] Value and Scope of SemClean-Bench.

A3. Thank you for raising this point. SemClean-Bench evaluates detection performance when obvious semantic shortcuts are less available, offering a complementary perspective on forensic perception. Besides, we conduct an extra visual comparison of the image and attention maps between the excluded and retained images in SemClean-Bench. As shown in Fig.5, one can see that the retained part has better visual quality and more diffuse attention patterns. On the other hand, the excluded part has conspicuous semantic anomalies that attract concentrated model attention, suggesting more dependence on obvious semantic shortcuts. We also revise the manuscript to clarify the benchmark’s evaluation scope.

[Q1] Transferability Across VLM Backbones.

A4. Thank you for raising this question. We extend Table 3(e) to InternVL3-8B, showing consistent improvements averaged across AIGI-Now, WildRF, and Chameleon.

BackboneBaseline (%)NFC (%)Gain
Qwen2.5-VL-3B66.994.1+27.2
Qwen3-VL-8B84.392.7+8.4
InternVL3-8B72.992.4+19.5

Although the optimal extraction layer differs across backbones, the same ablation-based strategy can be consistently adopted. We have added these results and clarifications into the revised manuscript.

[Q2] Clarification of Evaluation Differences

A5. Thank you for pointing this out. The reported results are measured on subsets filtered by removing generated images with obvious semantic anomalies, rather than on the original benchmarks. This makes detection more challenging and helps explain the lower scores compared with those reported on the original benchmarks.

[Q3] Performance on Real Images.

A6. Thank you for raising this point. For four existing AIGI benchmarks in SemClean, we added the corresponding real images following each benchmark’s original settings. On the other hand, four recent image generators in SemClean belong to text-to-image methods, thus having no matched real-image sets. We report overall accuracy, real accuracy and fake accuracy@FPR=1% in the following table, offering a fair comparison benchmark. NFC achieves higher average accuracy among all three metrics than Deep-VRM.

SourceNFC Overall Acc ↑NFC Real Acc ↑NFC Fake Acc @FPR=1% ↑Deep-VRM Overall Acc ↑Deep-VRM Real Acc ↑Deep-VRM Fake Acc @FPR=1% ↑
AIGI-Now99.599.999.995.599.794.3
Chameleon96.799.999.394.199.491.7
SynthBuster85.092.266.459.124.142.6
WildRF93.794.370.590.288.467.9
Avg.93.796.684.084.777.974.1
Reviewer_37fjOpenReview ↗

Response to authors

Dear authors,

Thanks for the detailed rebuttal. All of my concerns have been addressed.

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

The paper explores whether pretrained MLLMs already have useful forensic features that their usual vision-language pathway does not fully use. It proposes NFC, a connection that feeds intermediate features from a frozen vision encoder into the LLM. Training uses real–VAE reconstruction pairs first, followed by ordinary real/fake images. The paper also introduces SemClean-Bench, which contains generated images without obvious semantic errors. Results show strong detection performance, with ablations on the connection and training choices.

Critical strengths

The paper asks a useful question and proposes a simple approach. Using features already available in the frozen encoder could avoid the need for a separate forensic encoder or visual fine-tuning. I think this is worth exploring.

The results show that the connection helps. Table 3(a) reports an improvement from 95.1% to 97.2%, and Table 2 shows competitive results against Deep-VRM, especially on the Sem evaluation with matched LoRA settings. These provide encouraging evidence for the approach.

Critical weaknesses

Deep-VRM already injects forensic features into intermediate LLM layers (lines 134–141). NFC mainly differs in its use of frozen native features and its training procedure. Table 3(a) shows that adding the connection helps, but does not isolate the benefit of using native features. With explanation supervision, NFC and Deep-VRM achieve similar results (92.3% vs. 92.2%, Table 3(g)). Comparing feature sources under the same architecture and training settings would help clarify the contribution. Reporting training and inference costs would also show whether NFC offers a practical advantage.

SemClean contains only fake images and reports fake recall (lines 299–305), so a model that always predicts “fake” would score 100%. It is therefore unclear whether the gains reflect better discrimination or a stronger tendency to predict fake. The binary evaluations elsewhere partly address this concern, but not on the SemClean distribution. Adding matched real images and comparing recall at similar false-positive rates would make the results more convincing. Deep-VRM should also be included as a closely related baseline.

Figure 1 appears to favor visual layers around 10–15, whereas Table 3(f) favors layer 24. Please explain how this supports the claimed agreement with the probing results (lines 406–409). Table 3(c) also lacks a stage-2-only baseline, leaving the contribution of the first stage unclear. Clarifying the probing setup and adding this baseline would better support the design.

Questions for authors

Questions For Authors: See the “Critical Weaknesses” section. Weakness 2 is my main concern. I hope the authors can clarify it, as it raises a serious question about whether the benchmark allows a fair comparison between methods.

Other minor comments:

  1. Please provide the aligner architecture, LLM injection depth, and training settings. Lines 213–215 refer to the appendix, but these details are missing from the supplied version. Were the visual layer and mixing coefficient selected on a separate validation set? Table 3(a) reports 95.1% without NFC, while lines 364–367 report 95.5%. Please check this inconsistency.
2 comments

Dear Reviewer Aptn,

Thank you for the constructive review and for recognizing the motivation and empirical gains of NFC.

[W1] NFC’s contribution relative to Deep-VRM

A1. Thank you for your feedback. We address your comment in both feature selection and cost aspects.

(a) Feature source comparison under matched settings

We replace NFC’s native features with features from forensic encoder of Deep-VRM, using the same training settings. NFC yields higher accuracy (+1.5% average gain) on all three benchmarks, showing the benefit of using native features.

Injected featuresAIGI-NowWildRFChameleonAvg.
Deep-VRM98.491.297.495.7
NFC99.194.997.597.2

(b) Training and inference costs

We measure training GPU-hours, and inference FPS for both methods using the same environment. The results show that NFC achieves lower training costs and higher inference FPS than Deep-VRM, thanks to our simplified architecture.

Training GPU-hours ↓Inference FPS (images/s) ↑
Deep-VRM12.710.63
NFC5.390.81

[W2] SemClean recall and real-image false positives

A2. Thank you for raising this point. For four existing AIGI benchmarks in SemClean, we added the corresponding real images following each benchmark’s original settings. On the other hand, four recent image generators in SemClean belong to text-to-image methods, thus having no matched real-image sets. We report overall accuracy, real accuracy and fake accuracy@FPR=1% in the following table, offering a fair comparison benchmark. NFC achieves higher average accuracy among all three metrics than Deep-VRM.

SourceNFC Overall Acc ↑NFC Real Acc ↑NFC Fake Acc @FPR=1% ↑Deep-VRM Overall Acc ↑Deep-VRM Real Acc ↑Deep-VRM Fake Acc @FPR=1% ↑
AIGI-Now99.599.999.995.599.794.3
Chameleon96.799.999.394.199.491.7
SynthBuster85.092.266.459.124.142.6
WildRF93.794.370.590.288.467.9
Avg.93.796.684.084.777.974.1

[W3] Clarifying Layer Selection and the Contribution of Stage 1

A3. We thank the reviewer for the thoughtful comment. Below, We address your concern in two parts.

(a) Probing setup and interpretation of layer selection. Fig. 1 reports linear-probing results on real images and their paired VAE reconstructions, with peak performance around vision layers 10–15. In contrast, Table 3(f) evaluates NFC on real and generated images from three external public benchmarks and favors layer 24. Given these differences, we clarify that probing motivates our design, while the detection ablation determines the optimal layer.

(b) Contribution of Stage 1. We compare Stage 1 only, Stage 2 only, and both training orders, reporting the average and standard deviation over three runs with different random seeds. Our two-stage configuration consistently achieves the best average performance across AIGI-Now, WildRF, and Chameleon.

Stage 1 onlyStage 2 onlyStage 2 → Stage 1Stage 1 → Stage 2
Avg. ± SD (%)74.4 ± 2.394.7 ± 0.373.1 ± 1.997.1 ± 0.2

[Q1] Minor Comments.

A4. Thank you for pointing out these omissions and typos.

(1) We will add the aligner architecture and training settings in our manuscript, and explicitly specify the LLM injection index as 14, supported by the following ablation:

LLM injection index610141822
Avg. (%)94.896.297.296.696.1

(2) The visual layer and mixing coefficient were selected based on performance on the three external benchmarks, including AIGI-Now, WildRF and Chameleon, in Table 3(d) and (f).

(3) Sorry for the typo. The average reported in Table 3(a) was incorrect: the correct value is 95.5%, consistent with lines 364–367. We will correct the table accordingly.

Reviewer_AptnOpenReview ↗

Official Response

Thanks for the detailed rebuttal. All of my concerns have been addressed.

Reviewer_Nqa9OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whether low-level forensic representations must be learned by an additional forensic encoder or vision adaptation in MLLM-based AIGI Detection. The authors observe that native visual representations of pretrained MLLMs already have strong real/fake discriminability, but this information is underutilized in the standard vision-language pathway. To this end, this paper proposes Native Forensic Connection (NFC), which connects the middle layer features of the frozen vision encoder directly to the LLM through a lightweight aligner. And two-stage training is used to make effective use of forensic information. In addition, the authors construct SemClean-Bench to evaluate the detection ability of the model when semantic shortcuts are reduced by filtering obvious semantic anomalies. Experiments show that NFC achieves strong AIGI detection performance and generalization across generators without retraining the vision encoder.

Critical strengths
  1. Well motivated. The paper explores an interesting and meaningful perspective on MLLM-based AIGI forensics, shifting the focus from learning additional forensic representations to accessing existing native visual representations. The layer-wise probing and other observations provide meaningful motivation for this work.

  2. NFC is a simple yet effective approach that directly connects intermediate visual representations to the LLM without requiring additional forensic encoders or vision-side adaptation. Its design follows naturally from the identified representation access gap.

  3. The paper demonstrates competitive performance across diverse AIGI benchmarks and MLLM backbones. The probing analysis, ablation studies, and language-space evaluations provide supporting evidence for the effectiveness of NFC. SemClean-Bench further complements existing evaluations by reducing obvious semantic shortcuts.

Critical weaknesses
  1. The novelty of NFC need to be further clarified. NFC mainly relies on intermediate feature extraction and alignment, which are conceptually related to existing feature reuse techniques. Authors could better distinguish its scientific contribution from conventional intermediate feature injection approaches.

  2. The layer-wise probing results and extraction-depth ablations provide evidence that intermediate visual representations contain useful forensic information. However, their relationship with the actual performance gains of NFC could be further established through architecture-matched comparisons, such as an additional connection using final-layer representations under the same training settings.

  3. Although SemClean-Bench effectively reduces salient semantic anomalies, this does not necessarily establish reliance on low-level forensic evidence. The authors need to clarify its evaluation scope and avoid overstating its implications for low-level forensic perception. Additional implementation details on the observational studies and benchmark construction could be provided in the appendix to improve reproducibility.

Questions for authors
  1. Compared with existing frozen-backbone or intermediate feature injection methods, What are the key differences in research findings or method design of the representation access proposed by NFC? Can the authors further clarify the scientific contribution of this work?

  2. Could the authors provide an architecture-matched comparison using final-layer visual representations to further validate the benefit of accessing forensic-sensitive intermediate features?

  3. Could the authors further clarify the evaluation scope of SemClean-Bench, particularly how the reduction of semantic shortcuts relates to the assessment of low-level forensic perception?

1 comment

Dear Reviewer Nqa9,

Thank you for the constructive review and for recognizing the motivation and effectiveness of NFC.

[W1/Q1] Difference from existing feature injection techniques

A1. Thank you for the suggestion. We investigate other common feature injection methods, e.g., DeepStack (Bai et al., 2025), developed for general visual understanding, and compare it with NFC using the same settings. NFC improves the average accuracy from 89.6 to 92.7 across three benchmarks, showing the benefit of intermediate feature extraction and alignment through our proposed NFC. Besides, our NFC is motivated by the strong forensic discrimination of the native visual representations in pretrained MLLMs. In contrast, Deepstack utilizes multiple connections to jointly optimize the ViT and LLM for better visual understanding.

SettingAIGI-NowWildRFChameleonAvg.
DeepStack88.090.690.389.6
NFC91.197.190.092.7

[W2/Q2] Architecture-matched comparison using final-layer visual features

A2. We feed final-layer ViT features through the same NFC connection and compare them with intermediate features. Intermediate features achieve higher accuracy across all three benchmarks, suggesting that selecting them yields more effective representations for MLLM-based forgery detection.

SourceAIGI-NowWildRFChameleonAvg.
Final98.493.796.596.2
Intermediate99.194.997.597.2

[W3/Q3] Evaluation scope of SemClean-Bench

A3. Thank you for raising this point. SemClean-Bench evaluates detection performance when obvious semantic shortcuts are less available, offering a complementary perspective on forensic perception. Besides, we conduct an extra visual comparison of the image and attention maps between the excluded and retained images in SemClean-Bench. As shown in Fig.5, one can see that the retained part has better visual quality and more diffuse attention patterns. On the other hand, the excluded part has conspicuous semantic anomalies that attract concentrated model attention, suggesting more dependence on obvious semantic shortcuts. We also revise the manuscript to clarify the benchmark’s evaluation scope.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Review and Rebuttal Summary

Dear Area Chair and Reviewers,

We sincerely thank you for your time and constructive feedback. We appreciate the reviewers’ recognition of the following strengths.

  • Well-motivated research perspective. (Nqa9, Aptn, osLQ)
  • Simple and effective approach. (Nqa9, Aptn)
  • Competitive detection performance. (Nqa9, Aptn)
  • Comprehensive evaluation. (Nqa9, osLQ)
  • Clear presentation. (37fj)

We have updated the manuscript PDF with additional experiments, fixed typos and benchmark clarification. Below, we address each reviewer’s concerns and questions with the corresponding results and explanations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.