acceptodds
Under review as a conference paper at ICLR 2027

MMSec-R1: Test-Time Scaling for Fake Image Reasoning

Abstract

Current Multimodal Large Language Models (MLLMs) typically treat image forgery detection and localization as separate or loosely coupled tasks, often relying on massive supervised data. In this paper, we redefine this paradigm by proposing Fake Image Reasoning (FIR), a novel task that equates authenticity verification with the rigorous reasoning of forgery traces—essentially, an image is fake only if specific manipulated regions can be reasoned out. We introduce MMSec-R1, the first framework to explore Test-Time Scaling (TTS) for FIR via a multi-stage unified Reinforcement Fine-Tuning (RFT) strategy. Unlike traditional methods, MMSec-R1 treats precise localization as the definitive correctness verification, compelling the model to self-evolve thinking patterns from limited cold-start data. We validate our approach by training models across different scales (4B and 8B), both achieving superior performance. Our results demonstrate that scaling inference-time compute via RL effectively induces complex forensic reasoning in compact models, bridging the gap between detection confidence and visual evidence.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Fake Image Reasoning, Test Time Scaling

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 2 reviews, above 85% of the others.
Reviewer_Y4uqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MMSec-R1, a framework for Fake Image Reasoning (FIR) that jointly addresses image authenticity classification and manipulation localization. The authors first construct a small cold-start dataset of synthetic forensic chain-of-thought (CoT) demonstrations. They then apply reinforcement fine-tuning with DAPO, using a format reward and task-specific verifiable rewards, including a proposed Soft IoU score for bounding-box localization. A final DPO stage uses automatically judged preference pairs to improve consistency between generated rationales and final predictions. The approach is evaluated on in-domain and out-of-domain OpenMMSec splits and additional synthetic-image benchmarks. MMSec-R1-8B reports 90.34% average in-domain classification accuracy and 46.78 Soft IoU, together with 82.20% accuracy and 39.80 Soft IoU on the out-of-domain split.

Critical strengths
  1. Strong empirical performance on a broad forensic evaluation. The proposed training pipeline yields substantial improvements over the Qwen3-VL backbone on both classification and localization. 2.A useful empirical study of reinforcement fine-tuning components. The comparison among SFT-only, SFT+GRPO, SFT+DAPO, and the final DPO-aligned model helps identify where the reported gains arise.
Critical weaknesses
  1. Limited methodological novelty and an insufficiently supported emphasis on test-time scaling. Using ground-truth bounding boxes as outcome rewards for multimodal reinforcement learning is not, by itself, a novel idea, either for spatial grounding more broadly or for forgery localization. The paper needs a clearer differentiation from prior localization-rewarded visual reasoning and forensic methods. DAPO and DPO are also existing optimization methods rather than new contributions of this work. In Table 4, SFT+GRPO already attains 85.80% ACC / 42.00 Soft IoU, while replacing GRPO with DAPO yields 88.87% ACC / 43.67 Soft IoU: an incremental gain of 3.07 accuracy percentage points and 1.67 Soft IoU points. Although useful, this does not establish a fundamentally new training method. Moreover, the title makes test-time scaling the principal contribution, but Table 3 shows a relatively weak compute–performance trade-off. For MMSec-R1-8B, increasing the sampling budget from N=1 to N=8 raises latency from 1.05 s to 7.75 s (approximately 7.38×), while ACC increases only from 90.34% to 91.83% (+1.49 percentage points) and Soft IoU from 46.78 to 48.20 (+1.42 points). The paper does not sufficiently explain why this should be considered an important test-time scaling result rather than a modest benefit of repeated sampling. I suggest reframing the title and central contribution around localization-rewarded reinforcement fine-tuning unless much stronger compute–performance evidence can be supplied.

  2. The localization reward and generated CoT do not establish faithful visual reasoning; the reward formula itself requires correction or clarification. In DAPO, the reward primarily checks whether the predicted bounding boxes match the ground truth. A high localization reward does not verify whether the intermediate CoT correctly identifies the actual image artifacts, or whether the claimed forensic evidence is present at all. The later DPO stage targets logical consistency between textual reasoning and the final prediction, but such consistency is not equivalent to visual faithfulness. Consequently, the improved classification and grounding results do not establish the stronger claim that the model has learned reliable, causally grounded visual forensic reasoning. Eq. 3 appears unbounded as written because the two best-match IoU terms are summed, without normalizing by the numbers of ground-truth and predicted boxes. Let A = 1 denote the sum of best predicted-box IoUs for a single ground-truth box and B = 2 the sum of best ground-truth IoUs for two duplicate, perfectly overlapping predicted boxes. Then the stated expression yields Soft IoU = 2AB/(A+B+epsilon) ≈ 4/3 > 1. This raises concerns about duplicate predictions, false-positive penalties, reward hacking, and whether the implementation matches the formula. Since this is a core training reward and evaluation measure, it must be resolved. The cold-start CoT generation process explicitly provides the teacher model with ground-truth authenticity labels and, for locally manipulated images, ground-truth boxes (Section 3.1 and Appendix F). The teacher is then prompted to produce explanations invoking natural textures, lighting inconsistencies, and boundary artifacts. Such subtle forensic cues may not be reliably observable by a general-purpose VLM, especially at limited image resolution. The procedure could therefore produce post-hoc, label-conditioned rationalizations rather than evidence-based observations. The manuscript needs a more rigorous audit of teacher-generated explanations, and ideally should restrict them to visually verifiable cue types or report annotation/verification quality. Rule-based removal of conversational artifacts does not address this substantive hallucination risk.

Questions for authors

1.Prior work and novelty: Which specific aspect of the proposed bounding-box-rewarded RL formulation is new relative to existing spatial-grounding and image-forensic reinforcement learning approaches? Can the authors demonstrate a contribution beyond applying established DAPO/DPO recipes to an existing localization reward? 2.Soft IoU correctness: Is Eq. 3 identical to the metric and reward used in training and evaluation? How are multiple predicted boxes normalized, deduplicated, or penalized? 3.Test-time scaling protocol: How are the N independently sampled outputs in Table 3 combined or selected at inference time, when ground-truth labels and bounding boxes are unavailable? Can the authors report the trade-off against a fixed-compute baseline and disentangle gains from repeated sampling versus longer individual reasoning traces?

Reviewer_NtdbOpenReview ↗
Rating
4
Confidence
4(+1)
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces MMSec-R1, a framework that redefines image forgery detection and localization by shifting from simple pattern recognition to Fake Image Reasoning (FIR). To resolve the common disconnection between detection verdicts and visual evidence, the authors implement a multi stage Reinforcement Finetuning strategy using Task-Aware DAPO and Direct Preference Optimization to explore Test-Time Scaling in forensics. By utilizing precise localization as a definitive verifier, the approach enables compact models to self-evolve visual scrutiny capabilities from limited cold start data. The results demonstrate that scaling inference time compute allows models to bridge the gap between detection confidence and verifiable forensic traces, achieving SOTA performance across both in-domain and out-of-domain benchmarks.

Critical strengths
  1. A significant contribution is the transition from System 1 pattern matching to System 2 deliberate reasoning by treating precise localization as a definitive correctness verifier. Compared with FakeVLM or other SFT based models that merely mimic human textual styles, MMSec-R1 utilizes Reinforcement Fine-Tuning (RFT) with Soft IoU rewards to bridge the gap between linguistic logic and visual reality. The authors successfully validate that by rewarding spatial grounding, the model "self-evolves" internal monologues that are functionally coupled with the final detection verdict, effectively mitigating the decoupling and hallucination issues prevalent in previous imitation learning paradigms.

  2. The paper provides a compelling demonstration that forensic intelligence can emerge through scaling inference-time compute rather than solely increasing parameter counts. By employing the Task-Aware DAPO strategy, the framework stabilizes long-chain reasoning and allows compact 4B and 8B models to outperform significantly larger counterparts. The experiments on JPEG compression and Gaussian blur robustness further suggest that the model has not merely memorized high-frequency artifacts but has instead developed a robust reasoning policy anchored in structural inconsistencies, proving that structured reasoning is a more critical bottleneck in image forensics than model scale.

Critical weaknesses
  1. The core philosophy of the Fake Image Reasoning (FIR) paradigm is that an image is fake iff. verifiable forgery traces can be localized. However, Table 1 and Table 2 omit the Soft IoU metric for the AIGC category. Since AIGC involves global synthesis rather than local tampering, it lacks well defined bounding boxes to serve as definitive verifiers. This suggests the model's reasoning in AIGC tasks may lack the "verifiable spatial constraints" emphasized in the methodology, potentially causing the model to revert to pattern recognition (System 1) rather than the intended logical deduction (System 2) for this specific domain.

  2. While the paper introduces a DPO stage to align reasoning with final verdicts, the ablation study in Table 4 shows that MMSec-R1-8B-SFT (Cold Start only) actually performs worse than the base model in detection Accuracy. This indicates that the model's forensic capabilities rely almost entirely on the RL phase. Without a dedicated "Reasoning Quality" reward—only a Format Reward and Task-Aware Reward (Soft IoU). The model might learn to hack the Soft IoU by outputting boxes that overlap with ground truth while generating nonsensical or repetitive reasoning chains that do not truly drive the localization.

  3. The paper acknowledges in the Limitations that the impact of image encoder robustness across diverse pixel densities remains uninvestigated. This is a critical omission because forgery traces (especially in IMDL) are often resolution-dependent, and the efficacy of visual scrutiny may degrade significantly when the Test-Time Scaling is applied to low-resolution or heavily compressed images.

Questions for authors
  1. For weekness 1. In Table 1, why is the Soft IoU metric omitted for the AIGC task, and does the model generate bounding boxes for fully synthetic images? If the model cannot provide precise localization for global forgeries, how does the framework ensure that the generated CoT for AIGC remains anchored in visual evidence rather than hallucinating explanations to match a predicted label?

  2. For weekness 2. Since the reward function R primarily weights the Task-Aware Reward (Soft IoU) and Format Reward, what mechanism prevents the model from achieving high rewards through "Correct Box + Hallucinated Reasoning"? Furthermore, could the author provide a qualitative analysis or a specific metric to evaluate the logical necessity of the <think> block in reaching the final Soft IoU result?

  3. For weekness 3. The results in Table 3 demonstrate that increasing the sample budget N improves performance, but was this scaling effect consistent across different input resolutions? Specifically, does the model require higher inference-time compute to maintain localization precision when the visual forgery traces are attenuated by low resolution or high compression ratios?

I think this submission is really interesting. I read it in deep and came up with these concerns. I will keep my score and increase confidence after they are settled!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for submission 43948

Dear Acs, We thank Reviewers SE5m, Ntdb, and Y4uq for their constructive feedback. We summarize the main additions and the remaining disagreement below.

  1. Forensic supervision (Ntdb, Y4uq). We clarified why fully synthesized images (AIGC) receive image-level supervision and locally manipulated images (IMDL) receive box-level supervision. Adding pseudo-box supervision for AIGC increased false alarms and reduced localization accuracy on tampering tasks. Verifying individual local cues in fully synthesized images remains an open question.

  2. Reasoning and localization (SE5m, Ntdb, Y4uq). We clarified Soft IoU normalization and duplicate-box handling. In a new experiment with the model held fixed, shifting the region described in the rationale shifted the predicted box toward that region. Reviewer Ntdb maintained their acceptance score (4) and raised their confidence to 4 (3->4).

  3. Reporting corrections (SE5m). We clarified the training sequence and decoding settings and identified corrections to Tables 3 and 8. We also acknowledged that the reported human-audit percentages are inconsistent with the stated 300-sample protocol.

  4. Contribution of TTS (SE5m, Y4uq). Both reviewers questioned whether TTS accounts for enough of the overall improvement to warrant its prominence. Our evidence for its incremental contribution comes from holding the RL-trained 8B model fixed. Increasing the sample count from one to eight improved accuracy by 1.49 points and Soft IoU by 1.42 points on clean inputs at roughly 7× the latency. On low-resolution inputs, the gains reached 3.34 and 4.25 points, respectively. We believe that identifying when additional inference compute benefits an RL-trained forensic model is a major contribution of this work. We respectfully leave its significance to the Area Chair and Program Chairs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.