acceptodds
Under review as a conference paper at ICLR 2027

Awakening Verbalizer Forensics: Training-Free Readout Alignment for AI-Generated Image Detection in MLLMs

10 views

Abstract

Multimodal large language models (MLLMs) show promise for AI-generated image detection, but existing approaches mainly fine-tune them with binary real/fake supervision that, even with explanatory objectives, easily overfits to training fake patterns and fails to generalize. We find that frozen MLLM representations already contain transferable forensic information that the pretrained verbalizer, i.e., vocabulary projection, does not fully exploit. We introduce Forensic Readout Alignment (FRA), which estimates a forensic direction with a boundary offset using real images and single generator's fake images, and encodes them into the vocabulary projection weights of the real/fake answer tokens, without updating the backbone or any gradient-based training. FRA preserves the original distance and midpoint between the two answer rows, keeping the fitted boundary calibrated against the rest of the vocabulary. Using the same training data as the evaluated methods, FRA achieves the best overall cross-generator detection performance, outperforming LoRA-based adaptations, and remains effective across different training generators, training-set sizes, and MLLM backbones. Overall, FRA provides a lightweight way to effectively awaken pretrained forensic capabilities, while preserving the backbone's general language and visual understanding capabilities.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Deepfake Detection, Media Forensics, 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 3 reviews, above 85% of the others.
Reviewer_yydwOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a training-free Forensic Readout Alignment (FRA) method. Instead of updating the backbone of the multimodal large language model (MLLM), FRA only edits the lexical projection weights of real/fake answer tokens to mitigate the misalignment between the pretrained verbalizer and forensic discrimination directions. The forensic decision boundary is derived via Fisher direction estimation and carrier shift calibration. Experiments demonstrate state-of-the-art overall performance across seven cross-generator detection benchmarks, and the proposed method is compatible with multiple MLLM backbones. Finally, FRA retains the original question-answering capability of the model while achieving strong generalization for detection.

Critical strengths
  1. This work reveals that frozen MLLM representations already contain transferable forensic information, which cannot be fully exploited by the native lexical readout. It identifies the critical phenomenon of forensic verbalizer mismatch.
  2. A gradient-free FRA algorithm is proposed. It only modifies the weights of real and fake answer tokens and constructs the forensic decision boundary via closed-form estimation, without fine-tuning the backbone network.
  3. It achieves superior detection performance on multiple cross-generator benchmarks and across various MLLM backbones, with favorable robustness.
  4. The detection adaptation is decoupled from the native model capability. The modified MLLM fully preserves its original visual question answering generation ability, avoiding performance degradation induced by fine-tuning.
Critical weaknesses
  1. Although generalization experiments report promising results on multiple benchmarks, these results are obtained under the condition that the training set (especially fake data) covers a broad distribution. There lacks evaluations under fully non-overlapping data regimes. For instance, if the training set exclusively consists of GAN-generated samples from CNNDetection while the test set comprises diffusion-generated content (e.g., SD1.4, GLIDE, Midjourney, etc.), the projection of carriers may no longer approximate constants due to the substantial distribution shift between training and test domains, which could shift the decision boundary. This scenario has not been examined.
  2. Overhead: FRA requires pre-extraction of training-set hidden states for closed-form solving during inference. The additional engineering cost incurred during deployment should be discussed.

Overall, I consider this paper to feature a novel idea and sufficient technical contributions. Its merits outweigh its limitations, and I therefore assign a positive score.

Questions for authors
  1. The statement "2,500 class-balanced training images from GenImage (Zhu et al., 2023), with Stable Diffusion v1.4 (SD1.4) as the training generator" does not clearly specify the total volume and composition of the training dataset. Moreover, it remains ambiguous which test sets are fully disjoint from the training set, i.e., whether both real and fake test data have no distribution overlap with the training set.

  2. Why does GLM-4.6V-Flash in Table 7 yield 0% accuracy on EvalGEN, GPTImage2, MNW, and T2ICoReBench, which indicates complete label inversion? If so, simply mirroring the decision boundary should yield near-100% accuracy. However, the reported FRA only achieves an adjusted accuracy of 71.23% on GPTImage2.

  3. The current evaluation is limited to global binary detection of AI-generated images. For composite samples containing both authentic regions and AI-manipulated regions within a single image, how does FRA make predictions? Accordingly, the definition of whether a sample is regarded as real or AI-generated requires supplementary discussion.

3 comments
AuthorsOpenReview ↗

Response to Reviewer yydw — Part 1/3

We sincerely thank Reviewer yydw for the thoughtful questions and for recognizing our identification of the forensic verbalizer mismatch, the gradient-free readout edit, the robust cross-generator and cross-backbone gains, and the preserved original question-answering capability. We address the main concerns below.

Q1. Generalization and carrier stability under GAN-to-diffusion transfer.

R1. Thanks for this question. Our default fake training data come from SD1.4 alone, not a broad mixture (Section 4.1). Table 5 further fits FRA on each of eight GenImage generators separately, including GAN-based BigGAN (88.61% mean). We add a strict GAN-to-diffusion test (Table f):

  • GAN-to-diffusion setting. We fit FRA once on 1,250 real and 1,250 ProGAN images from CNNDetection. We then freeze it and apply it to 13 diffusion subsets from GenImage, AIGINow, and DRCT2M. Mean accuracy reaches 92.46%, with 91.92% real recall and 93.00% fake recall.
  • Carrier stability. The carrier satisfies v⊤h^≈1v^\top\hat{h} \approx 1, letting −bv-bv act as a constant offset −b-b. Under the shift its coefficient of variation rises from 0.18% (source) to at most 1.6% (targets), yet the mean projection stays within 0.012 of one for both classes on every subset. The offset thus remains near-constant and shifts both classes equally, so the boundary does not drift, which is consistent with the balanced recalls.

Table f: ProGAN-fitted FRA on unseen diffusion generators. Carrier = mean ± std of v⊤h^v^\top\hat{h}.

Benchmark / subsetReal carrierFake carrierReal recall (%)Fake recall (%)Acc (%)
GenImage / VQDM1.0025 ± 0.00521.0009 ± 0.004694.8097.6096.20
GenImage / GLIDE1.0023 ± 0.00551.0008 ± 0.005389.60100.0094.80
GenImage / ADM1.0023 ± 0.00601.0011 ± 0.005489.2096.4092.80
GenImage / SD1.41.0027 ± 0.00591.0024 ± 0.006992.8084.8088.80
GenImage / Wukong1.0030 ± 0.00551.0017 ± 0.007893.6082.8088.20
GenImage / Midjourney1.0031 ± 0.00561.0016 ± 0.008089.2092.4090.80
AIGINow / FLUX Pro (pixel)1.0027 ± 0.00521.0032 ± 0.010494.0089.0091.50
AIGINow / FLUX Krea (pixel)1.0028 ± 0.00591.0052 ± 0.009795.0086.0090.50
AIGINow / FLUX Dev (pixel)1.0020 ± 0.00600.9941 ± 0.015887.0093.0090.00
DRCT2M / LDM1.0019 ± 0.00521.0028 ± 0.005091.60100.0095.80
DRCT2M / SDXL Refiner1.0019 ± 0.00540.9884 ± 0.015494.4095.8095.10
DRCT2M / LCM-LoRA SD1.51.0020 ± 0.00591.0018 ± 0.010290.6097.0093.80
DRCT2M / SD-Turbo1.0020 ± 0.00511.0029 ± 0.010093.2094.2093.70
Mean––91.9293.0092.46

Q2. Separating one-time adaptation from online inference cost.

R2. Thanks for raising this point. We distinguish one-time adaptation from online inference, as summarized in Table g.

  • One-time adaptation. Other methods train with gradients over the listed epochs. FRA needs one forward pass plus a closed-form solve (68.3 + 0.4 = 68.7 s), with no gradients or epoch loop. All entries share the same 2,500 images and hardware, though the CLIP baselines adapt a far smaller backbone. Deployment loads only the modified weights and never revisits the data, features or solve.
  • Online inference. The architecture is unchanged: FRA runs at 81.70 ms/image, on par with the linear probe (82.84 ms) and far below LoRA + head (111.29 ms) and Deep-VRM (150.73 ms). Like any MLLM detector it still inherits the backbone cost.

FRA thus offers low-cost adaptation with no added inference overhead.

Table g: Adaptation and inference cost against existing methods.

MethodArch.EpochsAdaptation time (s)Infer time (ms per image)
UnivFDCLIP ViT-L/1417318.510.21 ± 0.02
FatFormerCLIP ViT-L/146421.124.39 ± 0.05
EffortCLIP ViT-L/1425480.514.81 ± 0.05
Deep-VRMQwen2.5-VL-7B2 + 41464.1150.73 ± 1.35
LoRA + headQwen3-VL-8B1491.6111.29 ± 0.08
Linear probeQwen3-VL-8B-71.282.84 ± 0.25
FRA (Feature extraction)Qwen3-VL-8B-68.3-
FRA (Solve + write-back)Qwen3-VL-8B-0.4-
FRA (Total)Qwen3-VL-8B-68.781.70 ± 0.16
AuthorsOpenReview ↗

Response to Reviewer yydw — Part 2/3

Q3. Training data volume and composition, and verification of test-set disjointness.

R3. Thanks for raising this question. Training uses 2,500 GenImage SD1.4 images (1,250 real ImageNet + 1,250 SD1.4 fakes; Section 4.1, Appendix A.3). The readout is solved from this set alone, with no test data in estimation or tuning. We verify independence at two levels in Table d.

  • Image level: zero overlap. All real test images come from COCO, Unsplash or other non-ImageNet sources, and every benchmark synthesizes its own fakes from its own prompts — including DRCT2M, whose SD1.4 images come from COCO captions.
  • Generator level: at most two residual overlaps. Only DRCT2M reuses SD1.4 among its 16 generators. Chameleon does not disclose the base models of its SD + LoRA images, so an SD1.x base cannot be excluded. The other five benchmarks are fully disjoint in both real and fake data.

Table h: Data sources and overlap analysis. "-": class absent.

Dataset#Real#FakeReal sourceFake generatorsImage overlapGenerator overlap
GenImage SD1.4 (training)1,2501,250ImageNetSD1.4--
AIGINow18,00018,000COCO (test)Nano Banana, GPT-4o, Jimeng, Kling, MiniMax, FLUX (Pro/Krea/Dev/Kontext)NoneNone
DRCT2M5,00080,000COCO 201716 SD-family models (LDM, SD1.4/1.5/2.1, SDXL, Turbo, LCM, ControlNet, inpainting)NoneSD1.4
EvalGEN-4,425-FLUX, GoT, Infinity, OmniGen, NOVANoneNone
GPTImage2-10,217-GPT-image-2NoneNone
MNW-16,250-Midjourney, DALL-E, GPT-image, Imagen, Firefly, Ideogram, Seedream, FLUX, SD3/SDXLNoneNone
T2ICoReBench-172,800-40 diffusion, AR, unified and closed-source models (FLUX, SD3.5, Qwen-Image, Infinity, Janus-Pro, BAGEL, GPT-Image, Nano Banana, Seedream)NoneNone
Chameleon14,86311,170UnsplashMidjourney, DALL-E 3, SD + LoRANoneUnverified

Q4. Interpreting GLM's near-zero scores and the limits of boundary mirroring.

R4. Thanks for pointing this out. The zero entries reflect a one-sided collapse of the original verbalizer, not an invertible boundary.

  • What 0% means. EvalGEN, GPTImage2, MNW and T2ICoReBench contain only fakes, so accuracy equals fake recall (Section 4.1): a near-zero value means the readout calls almost everything real, not that it separates classes with an inverted sign.
  • Why mirroring does not help. It merely turns a constant real predictor into a constant fake one, lifting fake-only subsets to 100% without any discriminative signal. On the three mixed-class benchmarks GLM sits at chance (50.03/50.01/50.01) and stays there after mirroring (49.97/49.99/49.99).
  • What FRA reports. The 71.23% on GPTImage2 is the fake recall of the source-fitted readout, with no target-specific mirroring or recalibration. The same fixed readout stays well above chance on all mixed-class benchmarks (75.89% Chameleon to 96.00% AIGINow).

Table i: GLM-4.6V-Flash on all seven benchmarks. Mirrored = 100 - Frozen.

BenchmarkMetric (%)FrozenMirroredFRA
AIGINowReal/fake ACC50.0349.9796.00
DRCT2MReal/fake ACC50.0149.9983.88
ChameleonReal/fake ACC50.0149.9975.89
EvalGENFake recall0.00100.0095.45
GPTImage2Fake recall0.00100.0071.23
MNWFake recall0.00100.0077.54
T2ICoReBenchFake recall0.0199.9997.73
AuthorsOpenReview ↗

Response to Reviewer yydw — Part 3/3

Q5. Behavior on partially manipulated images and the definition of the real/fake label.

R5. Thanks for this insightful comment. We clarify the label definition and FRA's prediction mode, and add an evaluation on locally edited images.

  • Label definition. Benchmarks label an image fake if any part of it was produced or altered by a generative model, regardless of the edited area. This underlies DRCT2M's inpainting variants and OpenSDID's local-editing splits, and we follow it.
  • How FRA predicts. The readout acts on the single final-layer hidden state of the whole image, so FRA gives an image-level verdict without localization: a partial edit is detected only if it shifts the global representation along the forensic direction.
  • Additional evaluation. On two class-balanced OpenSDID subsets of the OpenMMSec OOD split (FLUX.1 and SD family, 250 native real + 250 locally edited each, edits verified against the provided masks, no extra training or target calibration), FRA attains 75.40 mean accuracy, leading the best matched-training method by 16.1 points and the best released checkpoint (B-Free) by 7.9 points.

Table j: Locally edited image detection (Acc %), Matched (our 1,250 ImageNet real + 1,250 SD1.4 fake) / Released (official checkpoints).

MethodFLUXSD familyMean
CNNSpot53.80 / 50.0049.80 / 50.0051.80 / 50.00
UnivFD62.00 / 53.0056.60 / 52.2059.30 / 52.60
FatFormer59.40 / 50.0056.00 / 51.6057.70 / 50.80
AIDE52.20 / 50.4051.20 / 49.6051.70 / 50.00
DRCT52.60 / 51.2052.00 / 59.0052.30 / 55.10
Effort56.40 / 50.2053.40 / 51.6054.90 / 50.90
B-Free39.20 / 59.4042.00 / 75.6040.60 / 67.50
DDA53.40 / 55.8056.60 / 73.2055.00 / 64.50
Deep-VRM51.60 / 51.4050.00 / 50.2050.80 / 50.80
GenShield50.00 / -50.00 / -50.00 / -
FRA71.20 / -79.60 / -75.40 / -
Reviewer_bpD1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Forensic Readout Alignment (FRA), a training-free approach for AI-generated image detection using multimodal large language models (MLLMs). The authors argue that pretrained MLLMs may already encode transferable forensic information in their frozen representations, but their original vocabulary readouts fail to fully exploit this information. To address this issue, FRA estimates a regularized Fisher discriminant direction and calibrates the decision boundary using an offset carrier. The resulting readout is encoded into the vocabulary weights corresponding to the real and fake answer tokens while preserving their original midpoint and distance. The backbone remains frozen, and no gradient-based optimization is required. Experiments across seven benchmarks demonstrate substantial improvements over the original frozen MLLM and several existing detection methods. Further evaluations investigate generalization across different MLLM backbones and training generators, data efficiency, robustness to image perturbations, and preservation of ordinary visual question-answering behavior. The results suggest that readout-level alignment is a promising and computationally efficient alternative to conventional model fine-tuning.

Critical strengths
  1. A meaningful perspective on improving MLLM-based forensic detection. The paper identifies a potentially important distinction between the forensic information encoded in pretrained representations and the information actually exploited by the original output readout. Rather than assuming that improved detection necessarily requires additional representation learning, the authors investigate whether an appropriately aligned decision function can better utilize existing features. This perspective provides a useful alternative to increasingly complex fine-tuning and feature-integration approaches. 
  2. Strong empirical performance with a lightweight adaptation procedure. FRA achieves a reported mean accuracy of 91.10% across seven benchmarks, substantially outperforming the frozen Qwen3-VL-8B baseline at 66.38% and the evaluated LoRA variant at 74.40%. The method requires no gradient-optimized parameters and introduces negligible computational overhead beyond feature extraction. Experiments involving multiple MLLM backbones, training generators, and image perturbations further support the practical potential of the proposed approach.
Critical weaknesses

None

Questions for authors

1.Can the authors provide a more comprehensive comparison between FRA and conventional linear classifiers, including regularized logistic regression and explicit-bias variants, across all seven benchmarks?

  1. How do the authors verify the independence of training and evaluation images and establish that the target generators are genuinely unseen during fitting?
1 comment
AuthorsOpenReview ↗

Response to Reviewer bpD1

We sincerely thank Reviewer bpD1 for the constructive comments, and appreciate the recognition that FRA realigns the readout of frozen MLLM representations as a lightweight, training-free alternative to complex fine-tuning, performing strongly across backbones and training generators. We address the questions below.

Q1. Seven-benchmark comparison with conventional linear classifiers.

R1. We thank the reviewer for this suggestion. We extend Table 4 from three to all seven benchmarks and add a regularized logistic regression (L2 penalty with intercept), whose regularization strength is selected on source data only. All variants share the same frozen Qwen3-VL-8B features and the same 2,500 training images. FRA leads on every benchmark, improving the mean by 6.29 points over regularized LR, 7.37 over the linear probe, and 4.35 over the strongest explicit-bias variant. This suggests that, with limited source data, building on the pretrained readout transfers better than fitting a classifier from scratch.

Table c: Comparison with FRA and linear classifiers on frozen Qwen3-VL-8B features. All methods use the same features and training data.

MethodAIGINowDRCT2MEvalGENGPTImage2MNWT2ICoReBenchChameleonMean (%)
Frozen81.7268.9069.3242.4748.0679.1075.1166.38
Linear probe92.3183.1392.5970.1682.8993.0671.9883.73
Regularized LR92.1783.6294.0569.9683.2696.7873.8784.81
Fisher + bias87.7882.3794.1781.5784.0892.6567.3884.29
ww + Fisher + bias87.9283.1894.9582.0185.9794.5878.6486.75
FRA92.9486.7297.9586.1190.1599.5984.2691.10

Q2. Train–test data independence and unseen generators.

R2. Thanks for raising this point. The training set is 2,500 images from the GenImage SD1.4 subset, namely 1,250 real ImageNet images and 1,250 SD1.4 fakes (Section 4.1, Appendix A.3). The readout is estimated in closed form from this set alone, with no test data entering estimation or tuning. We verify independence at two levels in Table d.

  • Image level: zero overlap. All real test images come from COCO, Unsplash or other non-ImageNet sources, so none can coincide with a training image. Every benchmark also synthesizes its own fakes from its own prompts, including DRCT2M, whose SD1.4 images come from COCO captions and thus differ from the training images.
  • Generator level: at most two residual overlaps. Only DRCT2M reuses SD1.4 as one of its 16 generators, and Chameleon does not disclose the base models behind its SD + LoRA images, so an SD1.x base cannot be excluded. All other benchmarks, including AIGINow, EvalGEN, GPTImage2, MNW and T2ICoReBench, use generator families absent from the training set. To remove even these two cases, we re-estimate the readout once on CNNDetection (ProGAN fakes with real images) and evaluate the fixed readout on diffusion targets for which no test generator is seen during training. Mean accuracy reaches 92.46% over the 13 subsets in Table e.

Table d: Data sources and overlap analysis. "-": class absent.

Dataset#Real#FakeReal sourceFake generatorsImage overlapGenerator overlap
GenImage SD1.4 (training)1,2501,250ImageNetSD1.4--
AIGINow18,00018,000COCO (test)Nano Banana, GPT-4o, Jimeng, Kling, MiniMax, FLUX (Pro/Krea/Dev/Kontext)NoneNone
DRCT2M5,00080,000COCO 201716 SD-family models (LDM, SD1.4/1.5/2.1, SDXL, Turbo, LCM, ControlNet, inpainting)NoneSD1.4
EvalGEN-4,425-FLUX, GoT, Infinity, OmniGen, NOVANoneNone
GPTImage2-10,217-GPT-image-2NoneNone
MNW-16,250-Midjourney, DALL-E, GPT-image, Imagen, Firefly, Ideogram, Seedream, FLUX, SD3/SDXLNoneNone
T2ICoReBench-172,800-40 diffusion, AR, unified and closed-source models (FLUX, SD3.5, Qwen-Image, Infinity, Janus-Pro, BAGEL, GPT-Image, Nano Banana, Seedream)NoneNone
Chameleon14,86311,170UnsplashMidjourney, DALL-E 3, SD + LoRANoneUnverified

Table e: ProGAN-fitted readout transferred to diffusion subsets.

Benchmark / subsetACC (%)
GenImage / VQDM96.20
GenImage / GLIDE94.80
GenImage / ADM92.80
GenImage / SD1.488.80
GenImage / Wukong88.20
GenImage / Midjourney90.80
AIGINow / FLUX Pro(pixel)91.50
AIGINow / FLUX Krea(pixel)90.50
AIGINow / FLUX Dev(pixel)90.00
DRCT2M / LDM95.80
DRCT2M / SDXL Refiner95.10
DRCT2M / LCM-LoRA SD1.593.80
DRCT2M / SD-Turbo93.70
Mean92.46
Reviewer_Fo8ROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper finds that frozen MLLM representations already contain transferable forensic information that the pretrained verbalizer, i.e., vocabulary projection, does not fully exploit. Based on this observation, this paper proposes Forensic Readout Alignment (FRA), which estimates a forensic direction with a boundary offset and encodes them into the vocabulary projection weights of the real/fake answer tokens. FRA does not require updating the backbone or any gradient-based training. Extensive experiments with baseline demonstrate the effectivess of proposed method.

Critical strengths
  1. Interesting findings. The authors claim that they find that the frozen MLLM representations already contain transferable forensic information, which is an interesting insight overlooked by many previous MLLM-based methods.
  2. Efficient method design. The proposed FRA does not require any gradient-based training, which is highly efficient compared to other MLLM-based methods.
  3. Extensive experiments. The authors conduct extensive experiments, including generalization, ablations, visualization, and cost, which provide thorough support for the proposed method.
Critical weaknesses
  1. Deeper analysis of the reason behind. The authors find that MLLM representations already contain transferable forensic information that is not fully exploited by the pretrained verbalizer. But there is only a simple visualization in Fig.1 to support that, which I think is not enough. I also believe that MLLM contains much useful information that is not fully explored for AI-generated image detection, but it may not come from the verbalizer, the encoder, decoder, transformer, and middle layers may all contain useful information. Can the authors provide more theoretical proof or empirical evidence of their claim?
  2. Difference with traditional calibration. FRA uses labeled data to estimate a proper direction and modifies the output weights to obtain a more suitable boundary to distinguish real and fake images, which is very similar to traditional calibration in binary classification task, such as using different thresholds for classification or adding a classification head to the output layer. I believe all data distributions can be calibrated to fit a target distribution somehow. From this perspective, the novelty of FRA is somewhat limited. More justifications are needed.
  3. Overstated claim. The authors state that the proposed FRA is training-free and requires no trainable parameters, which I think is somewhat overstated since FRA still needs labeled data to estimate the direction and modifies output weights. More jusitifications are needed.
  4. Generalization of forensic detection. The direction is estimated using labeled data, which means the estimated direction is still highly related to the training patterns. Does this affect the generalization of FRA? Did the authors consider using different subsets to calculate the direction and compare their performance?
Questions for authors

Please see the weakness part above. My major concerns focus on the reason behind FRA, the difference with traditional calibration, and the generalization effect of different subsets used for determining direction. Thus, I currently lean towards nagative ratings. I hope the authors could address these issues properly during rebuttal.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Fo8R — Part 1/2

We sincerely thanks Reviewer Fo8R for the constructive comments, insightful questions. We greatly appreciate your recognition of our interesting finding, efficient method design, and extensive experimental evaluation. We also address the important concerns and questions raised below.

Q1. Source and evidence of the under-exploited forensic information.

R1. We appreciate the reviewer and agree that the vision encoder, the language-model layers and the intermediate layers may all contribute forensic information, but we do not claim it comes from the verbalizer. Our claim is that once forensic information reaches the frozen final representation, the native real/fake readout does not fully use it (Appendix A.1 explains the mechanism). Beyond Figure 1, the paper gives quantitative support:

  • Table 4. With fixed MLLM representation, mean accuracy rises from 73.31 to 89.34 with a standard linear probe and to 92.54 with FRA. Shifting only the boundary gives 75.06 (w/o direction). The bottleneck therefore lies in the readout direction.
  • Table 3. FRA improves all six MLLMs under fixed settings. On Qwen3.5, DeepSeek, Kimi, and GLM, the original readouts label almost all fakes as real and reach only 21–25%. With the same frozen representations, FRA raises them to 84–88%. Weak original detection performance thus reflects readout misalignment, not missing forensic information.
  • Table 8/10. With FRA, the full MLLM beats its vision encoder on GPTImage2 and Chameleon for all six backbones, while vision-only features are stronger on MNW for InternVL3.5, Kimi-VL and GLM-4.6V. Forensic information thus arises from both the vision encoder and the language-model layers, and the final representation aggregates both.

Q2. Difference and novety compared with traditional calibration.

R2. Thanks for your insightful comment. FRA uses a lightweight supervised fit to learn a decision boundary, a step common to supervised detection and calibration. Its contribution lies not in boundary fitting alone, but in integrating the fitted readout into the MLLM’s bias-free vocabulary head under explicit constraints: the carrier approximates a constant offset through existing weights (Section 3.3, Appendix A.2), and the write-back preserves the fitted boundary, the original row distance and the midpoint, the midpoint minimizing the weight change (Appendix A.4). Two distinctions:

  • Beyond threshold tuning. FRA realigns the readout as w′=w+η∥w∥u^−bvw'=w+\eta\|w\|\hat u-bv. Here, u^\hat u rotates it toward a forensic direction, and −bv-bv shifts the boundary. Threshold tuning keeps only the bb term, which gives 75.06 (Table 4), close to the frozen 73.31, versus 92.54 for FRA. The gain comes mainly from the direction term, which rotates rather than only shifts the boundary.
  • Beyond classification heads. An external head yields a label outside the vocabulary, so the verdict has to be injected back into the prompt or the decoded tokens before explanation. With FRA, the native real/fake token gives the verdict and directly conditions the explanation in the same autoregressive generation, while ordinary behavior is preserved (Table 7, Appendix A.11). With the same frozen features and training data, FRA outperforms the linear probe and both explicit-bias variants on all seven benchmarks (Table a). A plausible reason is that FRA corrects the pretrained readout instead of replacing it, which is more reliable than fitting a direction from scratch on 2,500 images.

Overall, FRA goes beyond recalibration. It realigns the readout direction, outperforms the evaluated traditional calibration, and encodes the detector into the MLLM's own verbalizer, natively bridging detection and explanation.

Table a: Comparison with linear probe and traditional calibration on Qwen3-VL-8B.

MethodAIGINowDRCT2MEvalGENGPTImage2MNWT2ICoReBenchChameleonMean (%)
Frozen81.7268.9069.3242.4748.0679.1075.1166.38
Linear probe92.3183.1392.5970.1682.8993.0671.9883.73
Fisher + bias87.7882.3794.1781.5784.0892.6567.3884.29
ww + Fisher + bias87.9283.1894.9582.0185.9794.5878.6486.75
FRA92.9486.7297.9586.1190.1599.5984.2691.10
AuthorsOpenReview ↗

Response to Reviewer Fo8R — Part 2/2

Q3. Clarifying "training-free" and "zero trainable parameters."

R3. Thanks for pointing out this ambiguity. We agree that FRA uses labeled source data and modifies the real/fake output rows, so we claim neither label-free nor weight-preserving.

  • Training-free. The term means FRA uses no backpropagation or iterative gradient optimization: direction, carrier and calibration coefficients are computed in closed form from labeled source data with fixed rules (Section 3), written into two vocabulary rows, backbone frozen.
  • Zero trainable parameters. This means zero gradient-optimized parameters (Table 1 caption), not that nothing is estimated.
  • Cost. A one-time feature extraction (68.3 s) plus solve and write-back (0.4 s), 68.7 s total, versus 318.5–1464.1 s for trained baselines (Table 6). We claim low-cost, not zero-cost, adaptation.

Q4. Generalization of FRA given a direction estimated from training data.

R4. Thanks for raising this point. We agree that a direction estimated from labeled source data depends on the source distribution. To verify generalization more thoroughly, we further evaluate FRA along two axes in our paper:

  • Different training subsets. Figure 3 varies the budget from 50 to 20,000 class-balanced images, refitting every method on five independent subsets per budget with fixed test sets (each subset is class-balanced with half ImageNet real and half SD1.4 fake). FRA is best at every budget, with std averaging 0.84 versus 1.41–14.81 for trained baselines. This stability is expected: frozen features already carry the signal, so only a boundary must be located, which is far less sample-sensitive than reshaping features.
  • Different source generators. Table b extends Table 5 to seven benchmarks and adds ProGAN. All nine source generators improve the seven-benchmark mean over the frozen reference (72.34–91.10 vs. 66.38). Diffusion sources close to the targets transfer best, whereas ADM, VQDM and ProGAN fall below frozen on GPTImage2, and ADM and VQDM also on Chameleon. FRA thus adapts across a range of source generators, although transfer quality depends on the source data.

Table b: Effect of the training source generator (one generator as sole fake source per row).

SourceAIGINowDRCT2MEvalGENGPTImage2MNWT2ICoReBenchChameleonMean (%)
Frozen81.7268.9069.3242.4748.0679.1075.1166.38
SD1.492.9486.7297.9586.1190.1599.5984.2691.10
SD1.593.5386.9297.7382.9789.9799.5182.1990.40
ADM88.1183.4790.6836.9959.2092.5571.6574.66
GLIDE85.9786.1595.6887.1878.2898.3276.1886.82
VQDM88.7584.4492.0529.5555.6989.7266.1572.34
Wukong93.5085.5997.0572.6081.2999.1082.6787.40
Midjourney94.2885.9997.7380.0490.7799.6477.2689.39
BigGAN88.9286.0887.2758.2465.4892.1687.8180.85
ProGAN83.9284.2188.8634.7471.7289.6979.6676.11

Comments on the paper

1 comment
Area_Chair_ssLoOpenReview ↗

Reminder: Please Review Author Rebuttals and Participate in Discussions

Dear Reviewers,

Thank you for your time on the reviews. The authors have now submitted their rebuttals, and I encourage you to read their responses and participate in the discussion.

Please consider whether the rebuttals address your concerns and, where clarification is still needed, engage with the authors through the discussion forum. I also encourage you to exchange perspectives with your fellow reviewers and update your review or rating if appropriate.

Your participation will help ensure a thorough and fair evaluation of each paper. Thank you again for your time and effort.

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.