acceptodds
Under review as a conference paper at ICLR 2027

Region-to-Fake: Perceptual Reinforcement learning for Fine-Grained Forgery Reasoning

Abstract

While Multimodal Large Language Models (MLLMs) significantly enhance forgery perceptual image understanding, they still exhibit limited perceptual reasoning capabilities when applied to the specialized domain of fine-grained image forgery detection. To address these issues, we propose the Region-to-Fake Distillation (R2F) framework. Our approach leverages large-scale MLLMs combined with image-cropping tool invocation to distill multi-round image crops and bounding box coordinates directly into the Chain-of-Thought (CoT). This enables the student model to acquire internalized bounding box localization capabilities, effectively eliminating the reliance on external tools during inference. To further optimize the model's perceptual understanding of forgeries while mitigating the prolonged inference latency, we introduce Perceptual-Aware Reinforcement Learning with Verifiable Rewards (PA-RLVR). By leveraging the intrinsic tag-generation capabilities acquired during the cold-start phase, this strategy enables direct importance sampling at the tag-sequence level, which significantly enhances the performance of forgery localization. Extensive experiments demonstrate that our method achieves high performance and efficient inference latency across multiple Fake Image Detection and Localization benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Perceptual Reinforcement learning, GRPO, Fake Image Detection and Localization

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_Yp1WOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work propose Region-to-Fake (R2F) for agentic-based image forgery detection. By firstly distill the knowledge CoT dataset, and secondly leverage the reasoning ability in agentic design, R2F has better performance on image forgery detection benchmarks.

Critical strengths
  • The proposed PA-RLVR is interesting, as designing the reward for forgery detection is diffucult.
  • The method IS-TS is explainable upon the design of reasoning trace. By using the <think> and <bbox 2d> tags, users could figure out how LLM and agents judge the suspecious image, instead of simply do the classfication.
  • Results among different benchmarks are competitive.
Critical weaknesses
  • Would the authors provide a detailed discussion on PA-RLVR, including the motivation on its design, and its difference to other related RL methods?
  • Would it be possible if the authors provide the comparison of close-sourced methods, like ChatGPT, on image forgery detection task?
Questions for authors

Please see my questions in Critical Weaknesses part.

Reviewer_KGFmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes Region-to-Fake (R2F), which improves multimodal LLMs' localization and detection of image forgery cues without external cropping tools at inference. Region descriptions, bounding boxes, and the final authenticity judgment are organized into tagged reasoning sequences, distilling tool-assisted region analysis into one model via SFT, with higher loss weights on box and answer tokens. PA-RLVR then applies reinforcement learning with tag-segment-level importance weights and rewards for reasoning, box IoU, and answer correctness. Using about 41K cold-start samples and 66K later trajectories, the paper evaluates on multiple authenticity benchmarks and reports localization on four datasets in the appendix. Experiments also cover image degradation, backbones, training data, and some loss/reward settings. R2F reportedly achieves detection accuracies of 90.14%, 79.02%, and 82.59% on LOKI, Chameleon, and OpenMMSec OOD.

Critical strengths

Coherent method design addressing a practical deployment problem:Distilling region coordinates and local evidence into one generation pass removes inference-time cropping tools, a clear practical motivation. The unified output format (description, localization, judgment) also makes the model's evidence easy to inspect. If localization quality is adequately validated, this would improve the usability of forensic models. Strong reported detection performance with some meaningful ablations:Table 1 covers multiple in-domain and OOD benchmarks. Table 5 shows TS-IS beating sequence sampling by 5.02 and 2.37 points on LOKI and Chameleon. Appendix Table 11 compares all-data SFT with staged RL, helping test whether gains come merely from more data. Despite the evaluation-consistency issues, these results support the direction.

Critical weaknesses

Key evaluations are internally inconsistent, and localization comparisons lack a clear unified protocol:In Table 4, Qwen3-VL-8B on LOKI has R-Acc, F-Acc, and Acc of 84.2, 87.2, and 59.8. Under standard binary statistics on one test set, overall accuracy is a weighted average of the class accuracies and cannot fall below both. Unless different subsets or aggregation were used and are stated, this row is erroneous. Also, Appendix H claims R2F has the best AUC on all four datasets, but Table 10 shows R2F's RTM AUC (0.580) below TruFor's (0.599). Moreover, Table 10 does not specify box vs. pixel-mask IoU, how other methods' outputs are converted, how multiple regions are aggregated, or how real images are handled, so fairness of the localization comparison cannot be confirmed. This directly affects the main “stronger detection and fine-grained localization” conclusion. The core PA-RLVR method is underspecified, and ablations do not directly validate its localization mechanism:Eqs. (3)–(5) define tag-segment weights but not how they enter the policy objective (advantage, clipping, KL). (R_{\mathrm{think}}) lacks an executable scoring rule, so “verifiable rewards” is unclear. Eq. (8) also lets every predicted box match the same reference box, so repeating one accurate box earns high reward without covering other regions. Tables 5 and 8 mainly report detection accuracy, not the contribution of TS-IS or box weighting to localization. The results thus do not adequately support the claimed improvement in fine-grained region perception.

Questions for authors

All items below are clarification or validation requests that could substantively change the assessment. Evaluation consistency and localization protocol:Please explain or correct the Table 4 LOKI metrics for Qwen3-VL-8B and the Appendix H RTM AUC claim. Please also specify Table 10's IoU definition, output conversion per method, multi-region matching, evaluation scope, and the source of image-level AUC scores. After correction, do the main conclusions still hold? Full RL implementation:Please provide the complete objective including TS-IS, and exact rules for (R_{\mathrm{think}}) and (R_{\mathrm{answer}}). How are rewards handled for real images, samples without reliable region annotations, and outputs with no valid box? Does Eq. (8) have unstated deduplication or coverage constraints? Attribution of gains and model selection:Under identical data and budgets, please report localization results for key ablations (with/without TS-IS, with/without box supervision/rewards). Was the final configuration in Tables 6 and 8 chosen on an independent validation set, so that OOD test independence can be assessed?

2 comments

As the rebuttal discussion period is approaching its deadline, we wanted to gently follow up to ensure our previous response has addressed all of your questions and concerns.

In our rebuttal, we provided the corrected Table 4 metrics, the detailed Table 10 evaluation protocol, the complete PA-RLVR policy objective and reward rules, and the component-wise localization IoU ablations under identical data budgets. Notably, regarding the shared questions on component-wise localization gains and the PA-RLVR formulation, Reviewer fvC3 confirmed that our additional experiments and clarifications have substantially addressed their concerns, and Reviewer Yp1W also expressed recognition of our PA-RLVR design and analyses.

We would greatly appreciate any feedback you might have before the discussion window closes, and we remain happy to provide any further clarifications if needed. If our responses have resolved your concerns, we would be very grateful if you could kindly reconsider your final rating.

Thank you very much for your time and thoughtful review!

Dear reviewer KGFm,

Thank you for your time and your constructive feedback on our paper. As the rebuttal period will end soon, we would greatly appreciate it if you could revisit our responses and let us know whether any concerns or questions remain. We hope these answers address the reviewer's concerns, and we would be happy to answer any further questions.

Thank you again for your time.

Reviewer_fvC3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Region-to-Fake (R2F), a framework for fine-grained forgery detection that aims to internalize region-level visual inspection into the reasoning process of multimodal large language models. The framework first constructs structured training trajectories containing bounding-box coordinates, regional observations, and authenticity predictions, distilling tool-assisted visual reasoning into a single MLLM through supervised fine-tuning. A tag-level weighted loss is introduced to emphasize localization and answer tokens during this stage.

The authors further propose Perceptual-Aware Reinforcement Learning with Verifiable Rewards (PA-RLVR), which incorporates Tag-Sequence Importance Sampling (TS-IS) to balance policy updates across different reasoning segments. The model is trained on multi-domain forensic data covering DeepFake, AIGC, image manipulation, and document forgery.

Experiments on several benchmarks, including LOKI, Chameleon, and OpenMMSec, demonstrate improvements in image-level detection accuracy and competitive localization performance. Additional ablations investigate the effects of tag-level weighting, reinforcement learning, and training data configurations.

Critical strengths

The paper addresses a relevant challenge in MLLM-based forensics: incorporating localized visual evidence into the reasoning process without relying on repeated external tool calls during inference. Distilling region-level observations and bounding-box predictions into structured reasoning trajectories is a reasonable approach that may improve the practicality of forensic MLLMs. The framework also considers multiple forensic domains, extending beyond conventional AI-generated image detection.

Critical weaknesses
  1. The reliability of region-level supervision and the localization reward is insufficiently established.

A central motivation of R2F is to improve fine-grained forensic perception through region-grounded reasoning. However, the training pipeline handles locally manipulated and fully AI-generated images differently. For locally manipulated images, bounding boxes can be derived from existing localization annotations, whereas fully generated images without ground-truth boxes rely on teacher-generated suspicious regions.

For these samples, the paper describes filtering based primarily on answer correctness and output format, but does not clearly establish the reliability of the generated bounding boxes. More importantly, PA-RLVR defines its localization reward using IoU against ground-truth boxes, yet it remains unclear how this reward is computed for samples without reliable localization annotations.

This distinction is important because teacher-selected suspicious regions are not necessarily equivalent to objectively verified forgery regions. Without further clarification or validation, it is difficult to determine whether the model learns reliable forensic grounding or primarily reproduces teacher-generated region descriptions.

  1. The experimental comparisons and task-specific validation do not fully support the breadth of the claims.

The paper presents R2F as a unified multi-domain forensic MLLM, but the baseline comparisons are somewhat inconsistent across evaluations. Table 1 primarily compares against general-purpose and forensic MLLMs, whereas Table 2 evaluates the four OpenMMSec domains against conventional vision-based detectors without including existing unified forensic MLLMs. This makes it difficult to assess the relative advantages of R2F under a consistent multi-domain MLLM evaluation setting.

Furthermore, although fine-grained localization is central to the motivation, most main experiments and component-wise ablations focus on image-level detection accuracy. Localization results are mainly reported in the appendix, without sufficiently isolating the contributions of the proposed distillation and reinforcement learning strategies to localization quality.

The methodological novelty is also somewhat incremental, as tool-use distillation and structured loss weighting have established precedents, while TS-IS primarily modifies the importance sampling granularity of existing policy optimization methods. Stronger localization-oriented validation would help substantiate the significance of this adaptation.

Questions for authors

Bounding-box supervision and verifiable rewards: For fully AI-generated images without ground-truth bounding boxes, how are teacher-generated suspicious regions validated? During PA-RLVR, are these pseudo boxes treated as ground truth for the IoU-based reward, or is the localization reward disabled for such samples? How does the method prevent unreliable pseudo-localization from introducing noisy supervision?

Multi-domain baseline fairness and completeness: The baseline selection and training data configurations appear inconsistent across different evaluations. In Table 1, R2F is compared against existing MLLM-based detectors on AIGI benchmarks, although these methods may use substantially different training data and domain coverage. In contrast, Table 2 evaluates R2F across the four OpenMMSec domains primarily against conventional vision-based detectors, rather than comparable unified forensic models. Notably, while R2F builds upon the OpenMMSec/SICA data framework, the unified detection method proposed in SICA is not included as a baseline. Could the authors clarify the training data differences across the compared methods and explain the baseline selection criteria? Including the SICA baseline and, where feasible, data-aligned comparisons would help establish whether the reported improvements stem from the proposed training strategy rather than differences in training data composition or coverage.

Contribution to localization: Could the authors provide a component-wise localization evaluation comparing standard SFT, Tag-Loss, GRPO, and TS-IS? This would help determine whether the proposed optimization substantially improves fine-grained localization rather than primarily image-level classification accuracy.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.