acceptodds
Under review as a conference paper at ICLR 2027

Know What You Don't Know: Confidence-Calibrated MLLMs for Tampering Text Detection

Abstract

Multimodal large language models (MLLMs) can detect, localize, and explain text-image tampering, but lack explicit confidence outputs, which weakens the reliability of detection systems in practical scenarios. While calibration methods allow models to output confidence matching prediction accuracy, they optimize confidence separately and only consider correctness in calibration. Under such paradigm, models cannot learn forensically beneficial knowledge from calibration, nor can they perceive the certainty of their own analytical process. To address this, we propose a three-stage confidence-calibrated training framework that connects self-assessment with forensic learning. First, forensic continual pretraining on large-scale text-image data develops task-specific perceptual knowledge. Second, supervised fine-tuning introduces structured reasoning, answer, and confidence outputs together with a **Confidence Grounding Attention Module** (CGAM). CGAM uses the confidence representation to aggregate selected reasoning and answer representations, with auxiliary authenticity supervision tying this pathway to forensic prediction. Third, Group Relative Policy Optimization jointly optimizes forensic rewards and a **Forensic-Quality-Aware Calibration** (FQAC) reward. The calibration target combines current-rollout correctness and cross-rollout accuracy with token-level reasoning certainty, using the latter as an auxiliary process signal rather than a substitute for empirical correctness. Experiments on four test sets show that our framework outperforms TextShield-R1 and achieves an average Expected Calibration Error of 21.4%, lower than the evaluated calibration baselines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Tampering Text Detection, Confidence Calibration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_sc9tOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the lack of explicit confidence in MLLM-based tampered text detection. On a Qwen3-VL-8B backbone, it proposes a three-stage confidence-calibrated training framework that connects self-assessment with forensic learning.

Critical strengths

S1: This paper targets a real gap in forensic MLLMs: they output verdicts but no confidence that could decide between automatic acceptance and manual review.

S2: This paper evaluates all models without OCR Rectification.

Critical weaknesses

W1: This paper does not show that "lucky guesses" occur, and Sec. 3.4 admits its remedy, EthinkE_{think}, does not measure reasoning correctness. CGAM supervises class labels rather than correctness, with no theoretical justification. This paper never isolates each stage.

W2: This paper is incremental over TextShield-R1, reusing its rewards, data split and annotations. CGAM attends over only two tokens, and replacing hconfh_{conf} with a learnable query changes ECE by only 0.7 (Table 4). This paper omits the closest recent baselines: MLLM text-tampering methods (TextSleuth, Xu et al. ACM MM 2025, DocShield) and calibration methods (VL-Calibration, temperature scaling). It reports only ECE, without Brier, AUROC or risk–coverage results.

Questions for authors

Q1: Table 3, row 1 has no <conf> supervision yet reports an ECE of 36.8. How is its confidence obtained?

Q2: Can the authors report SFT+RL without forensic pretraining, and the full model without CGAM? Also, why does SFT on 25% of the data (Table 3, 89.3 Cls) exceed full-data SFT (Table 7, 87.9)?

Q3: Under teacher forcing, is the <conf> label computed on the TFR-annotated <think> or on a sampled one? Without a format reward, how are malformed outputs handled, and what is the parse-failure rate?

4 comments
AuthorsOpenReview ↗

Response to Reviewer sc9t (Part 1)

Response to W1: Lucky Guesses, CGAM Label Design, and Insufficient Isolation of Training Stages

We thank the reviewer for this critical comment.

(1) Lucky guesses do occur. Concretely, on the four test sets 7,613 of Stage-2's 19,869 tampered/generated predictions are "correct-but-flawed" predictions—the image is classified correctly, yet its forensic analysis is wrong in at least one respect (localization error, OCR error, or an evidence-inconsistent explanation). This amounts to 38.3% of all tampered/generated samples, confirming that lucky guesses do occur.

(2) The scope of EthinkE_{\text{think}}. EthinkE_{\text{think}} is an auxiliary signal about the certainty of the model's own reasoning trace. Through the CGAM-guided aggregation in Stage 2, optimizing it via hconfh_{\text{conf}} endows it with the ability to express the correctness of the forensic analysis process; in Stage 3 it is fused with AcrossA_{\text{cross}} and AcurA_{\text{cur}} to capture process and outcome correctness, respectively.

  • As an empirical check (Table 2 (W1-2)), we take Qwen3-VL-32B-Instruct with privileged ground-truth forensic knowledge as a reference and correlate its teacher-forced per-token distribution with our 8B model's over the same <think> sequence (Pearson ρ\rho). As EthinkE_{\text{think}} rises, ρ\rho grows monotonically from 0.310 to 0.508 (+0.198) while accuracy climbs from 88.2% to 98.0%, indicating that EthinkE_{\text{think}} tracks agreement with a correct analysis.

(3) CGAM's supervision target and its design rationale. hconfh_{\text{conf}} needs to express the process correctness of the forensic analysis, whereas htsh_{\text{ts}} and hansh_{\text{ans}} need to express image authenticity; correctness itself carries no information about image authenticity and is therefore unsuitable as a supervision label. The image class label both expresses authenticity and is a necessary condition for verifying process correctness, so we use class labels as the auxiliary supervision for CGAM.

(4) Isolation of training stages. SFT and GRPO isolation is covered by Table 3 of the main text, and the Pre-train-only model is reported in Table 7 of the appendix. We also report the SFT+GRPO-only configuration to show the role of the Pre-train stage more completely, consolidating all results in Table 3 (W1-5) and Sec. 4.4 of the revised manuscript.

We will include the above experiments and analyses in the revised manuscript.

Table 2 (W1-2): ρ\rho and classification accuracy across EthinkE_{\text{think}} quintiles

Mean EthinkE_{\text{think}}Mean ρ\rho↑Cls. Acc (%, ↑)
0.62080.31088.24
0.62640.36490.20
0.67140.40288.24
0.71840.44491.18
0.72820.50898.04

Table 3 (W1-5): Stage isolation (average over four test sets)

ConfigurationPre-train (Stage1)SFT (Stage2)<conf> (Stage2)CGAM (Stage2)RforR_{\text{for}} (Stage3)RconfR_{\text{conf}} (Stage3)Cls.OCRLoc.Res.ECE↓
Pre-train✓–-–––38.8–40.9––
Pre-train + SFT✓✓-–––89.354.650.552.736.8
Pre-train + SFT + <conf>✓✓✓–––89.652.347.951.726.7
Pre-train + SFT + <conf> + CGAM✓✓✓✓––90.855.450.855.424.0
Pre-train + SFT + <conf> + CGAM + RforR_{\text{for}}✓✓✓✓✓–91.556.551.256.923.8
SFT + <conf> + CGAM + RforR_{\text{for}} + RconfR_{\text{conf}} (new)-✓✓✓✓✓70.619.618.926.039.7
Pre-train + SFT + <conf> + RforR_{\text{for}} + RconfR_{\text{conf}} (new)✓✓✓-✓✓90.255.851.054.622.3
Pre-train + SFT + <conf> + CGAM + RforR_{\text{for}} + RconfR_{\text{conf}}✓✓✓✓✓✓91.056.851.757.421.4
AuthorsOpenReview ↗

Response to Reviewer sc9t (Part 2)

Response to W2: Is the Method Merely an Incremental Improvement over TextShield-R1?

We thank the reviewer for the candid assessment of our novelty and experimental completeness.

(1) Methodological differences from TextShield-R1.

Reusing TFR's data split and forensic rewards is a controlled variable: it makes the way confidence enters the learning process the only variable that changes, rather than inheriting TextShield-R1's methodological setup.

The core innovation of this work relative to TextShield-R1 is:

  • Methodologically, we are the first to propose that a calibration signal is not merely a scalar aligned with accuracy; it can serve as a learning signal for the forensic optimization process. As Table 3 of the main text shows, simply adding a confidence output on top of a forensic model improves calibration at the expense of forensic performance (ECE decreases by 10.1%, while OCR/Loc./Res. drop by 2.3%/2.6%/1.0%). With CGAM, designed according to our methodology, the confidence semantic representation participates in learning the forensic representation as an aggregation condition, and this trade-off disappears (ECE decreases by 2.7%, while Cls./OCR/Loc./Res. improve by 1.2%/3.1%/2.9%/3.7%). Hence, our methodological innovation is not the addition of a confidence output, but making the confidence signal an effective signal for forensic optimization.

  • In terms of practical value, our method can be used to build automated business workflows: when the model is uncertain about a prediction, the sample is routed to manual review, reducing the risk of misjudgment. Beyond the ECE reported in the paper, we further analyzed the ECE of the auto-accepted region: on the four test datasets, we retain only samples with C^≥0.9\hat{C}\geq 0.9 as the auto-accepted set (results that can be accepted directly in an automated workflow) and recompute ECE within this set using the same 10-bin binning scheme. The results show that this threshold automatically accepts 48.0% of the test samples, while the ECE of the auto-accepted region is only 7.31%, far below the full-set value of 21.4%. In other words, our method can cut the manual review workload by nearly half, with a low risk of misjudgment for auto-accepted samples—an advantage that rests precisely on accurate confidence outputs.

In summary, the proposed method is not incremental over TextShield-R1, but a complete innovation built on a new methodology that can be deployed in practice.

(2) CGAM uses only two tokens. We emphasize that this is a deliberate choice, not a capability limitation: htsh_{\text{ts}} is the initial reasoning state and hansh_{\text{ans}} the conclusion state, corresponding respectively to the starting point of analysis and the endpoint of the verdict—the two most representative states in the forensic decision chain. We do not claim that the 0.7 ECE difference in Table 4 of the paper is a significant improvement; more important are the simultaneous gains in the forensic metrics (+0.4/+1.9/+1.0/+1.3), which show that CGAM achieves joint forensic–calibration optimization rather than calibration alone.

(3) Missing the closest baselines. TextSleuth and DocShield are not yet open-sourced. Due to time constraints, we will report the replication forensic and calibration metrics in the final revised version.

(4) Reporting only ECE. We accept this comment and have added Brier Score and AUROC (Table 4 (W2-4) below), as well as Risk–Coverage (Table 5 (W2-5) below).

Table 4 (W2-4): Comparison of calibration baselines (ECE from Table 2 of the paper; all other results are new)

MethodECE↓Brier↓AUROC↑
Baseline26.718.341.0
RLCR23.716.147.8
DCPO22.916.147.9
Ours21.415.255.0

Table 5 (W2-5): Risk–Coverage of Selective Prediction (Risk↓ in %, average over four test sets)

CoverageBaselineRLCRDCPOOurs
100%10.419.319.769.07
90%10.889.4310.049.25
80%11.639.8110.559.35
70%12.6310.0910.608.81
60%13.8310.7410.718.48
50%12.1810.3410.237.55
AuthorsOpenReview ↗

Response to Reviewer sc9t (Part 3)

Response to Q1: How is the ECE of 36.8 in the First Row of the Paper's Table 3 Obtained Without <conf> Supervision?

We thank the reviewer for pointing out this implementation detail. We acknowledge that the original manuscript did not explain it sufficiently. The actual situation is: the confidence in the first row comes from the inference prompt's direct request to the model.

Specifically:

  • All configurations in this paper share the same inference prompt (Figure 6 of the main text), which explicitly asks the model to give a confidence in the 0–1 range inside the <conf> tag. Thus “no <conf> supervision” means that the confidence value is not used as a supervision target during training, not that no confidence is produced at inference time.
  • The model in the first row learns only the content and format of <think>/<answer> during SFT; the <conf> field has no label constraint. The value it outputs is purely prompt-induced verbalized confidence, which simulates a forensic detector that has not undergone any calibration (confidence) training. This is precisely the starting point our work aims to improve.

Revision: Sec. 4.4 will add a sentence explaining the source of the confidence in the first row (a prompt-induced verbalized output).

Response to Q2: SFT+RL Without Pre-train, the Full Model Without CGAM, and Why 25% SFT Outperforms Full-Data SFT?

We thank the reviewer for raising these issues. We have added more complete ablation results in Sec. 4.4. The two newly added configurations are tabulated in Table 3 (W1-5) of our response to W1.

(1) SFT + RL without Pre-train. Added: starting from the original Qwen3-VL-8B, skipping the first-stage forensic continued pre-training, and directly performing the same SFT (25% TFR) and GRPO training (same rewards, same number of steps). Given the challenge of joint forensic–calibration optimization, the Pre-train stage is crucial for building the model's basic forensic capability.

(2) Full model without CGAM. Added: we keep Pre-train, SFT, and GRPO (including RforR_{\text{for}} and RconfR_{\text{conf}}) and remove only CGAM and Lcls\mathcal{L}_{\text{cls}}. The no-CGAM row gives 90.2/55.8/51.0/54.6 and ECE = 22.3 (full model: 91.0/56.8/51.7/57.4, ECE = 21.4); both the forensic and calibration metrics degrade to some extent.

(3) Why 25% SFT outperforms full-data SFT. The cause is overfitting: the 25% SFT baseline is trained for 1,212 optimizer updates, whereas a single epoch over the 100% pool already reaches 1,567, so full-data SFT keeps optimizing well past the 25% budget and degrades on held-out data. We re-ran Stage 2 from the same Pre-train weight on the 100% pool with the identical recipe (CGAM removed) and evaluated checkpoints on the fixed 1,000-sample subset (Table 6 (Q2) below).

Table 6 (Q2): Full-data SFT as a function of optimizer updates (fixed 1,000-sample subset; four-test-set macro)

ConfigurationSFT dataUpdatesCls.↑OCR↑Loc.↑Res.↑ECE↓Brier↓
25% SFT + <conf>, two epochs25%1,21290.1051.4146.9952.0026.580.1798
100% SFT + <conf>, matched updates100%1,20090.2053.8048.9653.5224.770.1691
100% SFT + <conf>, mid-training100%1,50088.3052.7047.4250.5126.900.1888
100% SFT + <conf>, one epoch100%1,56787.5050.8646.3350.1026.670.1933

Two observations confirm this:

  • At a matched update budget, more data is harmless (even slightly better). At ≈1,200 updates the 100% run beats 25% (Cls. 90.20 vs. 90.10, ECE 24.77 vs. 26.58); the two splits also share the same source subsets and class proportions, so the extra examples add few new patterns.
  • The degradation comes from the extra updates. From step 1,200 to one epoch, held-out Cls. falls 90.20 → 87.50 and ECE rises 24.77 → 26.67 while training loss keeps decreasing (0.0490 → 0.0478)—a clear train/test divergence. The paper's 2-epoch 100% SFT is the same over-optimized regime, which our full four-test-set run reproduces (87.56/52.99/49.35/49.85, ECE 26.83 vs. paper's Table 7 row 2 87.9/53.2/48.9/50.3, ECE 26.3).
AuthorsOpenReview ↗

Response to Reviewer sc9t (Part 4)

Response to Q3: How Are <conf> Labels Generated Under Teacher Forcing, and How Are Format Errors Handled?

We thank the reviewer for raising this reproducibility issue.

(1) How <conf> labels are generated:

At each training step of Stage 2, the <conf> label is produced by a gradient-free forward pass with the current parameters over the entire TFR-annotated teacher-forced <think>…<answer>… sequence. We take the mean per-token cross-entropy ℓˉ\bar\ell over the annotated <think> span to obtain c=clip(e−ℓˉ,0,1)c=\mathrm{clip}(e^{-\bar\ell},0,1), then insert the tokens of <conf>c</conf> before the eos of that sequence and write them into the labels.

(2) How the format is enforced: the implementation includes a separate Format reward. The Format reward uses a regular expression to check whether the output strictly matches <think>…</think><answer>…</answer><conf>…</conf>, assigning 1 if valid and 0 otherwise. The revised manuscript will add a note on this implementation detail below Eq. (8); we thank the reviewer for pointing it out.

(3) Parse-failure rate: it has been measured and reported, and failed samples are not discarded. Table 7 (Q3) reports the format-error rates at each training stage (missing <think>/<answer>, missing or unparseable <conf>, and the overall failure rate), covering the post-SFT and post-GRPO checkpoints, with the Baseline as a reference.

Supporting Material

Table 7 (Q3): Parse-failure rates and two evaluation protocols

CheckpointMissing <think>Missing <answer>Missing or Unparseable <conf>Overall Parse-Failure Rate
Baseline (no <conf> supervision)0.0076%1.1706%15.5643%16.7273%
After SFT0.0025%0.0025%0.0025%0.0025%
After GRPO (full model)0.0051%0.0051%0.0051%0.0051%
Reviewer_LxEUOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies confidence calibration for multimodal large language models in text-image tampering detection, aiming to improve both forensic performance and the reliability of reported confidence. It proposes a three-stage framework comprising forensic continual pretraining, supervised fine-tuning, and reinforcement learning. The Confidence Grounding Attention Module (CGAM) connects confidence-related representations with forensic learning through an auxiliary authenticity classification objective. The Forensic-Quality-Aware Calibration (FQAC) reward combines reasoning-token probabilities, current-response correctness, and correctness across sampled responses to construct a confidence target.

Critical strengths

None

Critical weaknesses

Weaknesses

  • The methodological novelty could be more clearly established. The framework builds on existing ideas in forensic training and confidence calibration. Although the combination is potentially useful, its distinction from closely related approaches and the need for task-specific design are not sufficiently clear.

  • The added complexity is not fully justified. The framework introduces several training stages and components. While the ablations provide useful evidence, they leave some uncertainty about which elements drive the improvements and whether a simpler approach could achieve comparable results.

  • The interpretation and usefulness of the confidence signal need stronger support. Reasoning-token probability may capture factors other than the reliability of visual analysis, as the authors acknowledge. The reported calibration results are encouraging, but provide limited evidence about the signal’s role and its usefulness for identifying unreliable predictions.

Questions for authors

Questions

  • Could the authors clarify the main methodological distinction from existing calibration approaches and why it is particularly suited to text-image tampering detection?
  • What do the current experiments suggest about the most important components, and whether the full framework is necessary?
  • Could the authors provide further evidence or analysis clarifying what the reasoning-confidence signal captures and how it helps identify unreliable predictions?
2 comments
AuthorsOpenReview ↗

Response to Reviewer LxEU (Part 1)

Response to Q1: What Is the Main Methodological Innovation Compared with Existing Confidence Calibration Methods?

We thank the reviewer for pointing out that the methodological novelty needs to be clarified further. Unlike existing calibration methods, which typically calibrate only against prediction correctness, this paper proposes two designs tailored to text-image tampering detection:

  1. CGAM: using the confidence representation as the Query, it conditionally aggregates the reasoning and answer representations and, through auxiliary authenticity-classification supervision, establishes an explicit link between the confidence representation and forensic discriminative information—the calibration signal can therefore feed back into the forensic representation, so confidence is no longer an independently optimized target.
  2. FQAC: it incorporates current-prediction correctness, cross-rollout correctness, and the Token Probability of the thinking process into the confidence target—introducing correctness about the forensic analysis process and preventing accidentally correct guesses from being assigned high confidence.

These two designs address the specific requirements of text-image forensics:

  • The model must not only decide whether a text-image has been tampered with, but also provide fine-grained information such as the tampered location, the tampered text, and the manipulation type. Calibration based solely on the final classification result cannot exploit information from the forensic analysis process. We therefore propose CGAM, which uses confidence semantics as one of the conditions of forensic optimization, achieving joint forensic–calibration optimization.
  • As a highly structured task (binary authenticity classification), the model has a high probability of obtaining the correct classification by random guessing. FQAC therefore introduces the Token Probability of the thinking process: building on Stage 2, it brings the captured correctness of the model's forensic analysis process into calibration training (Stage 3).

Response to Q2: Which Modules Actually Bring Performance Gains? Is the Full Framework Necessary?

We thank the reviewer for the suggestion. The results in Table 3 of the main text already show:

  • Adding CGAM: all four forensic metrics improve (90.8/55.4/50.8/55.4), and ECE further drops to 24.0;
  • Adding RforR_{\text{for}}: the forensic metrics continue to improve (91.5/56.5/51.2/56.9), with little change in ECE;
  • Adding RconfR_{\text{conf}}: ECE reaches its lowest value of 21.4, with little change in Cls./OCR/Loc./Res.

In addition, we have added two key control experiments:

  • No Pre-train: starting directly from the original Qwen3-VL-8B for SFT + GRPO;
  • Full three-stage training, but with CGAM removed from Stage 2.

The results in Table 1 (Q2) below show that after removing Pre-train, all forensic metrics decline and ECE rises to 39.7%. Given the challenge of joint forensic–calibration optimization, this indicates that Pre-train—which equips a general-purpose MLLM with basic forensic capability—is crucial for preserving the model's ability to extract fine-grained forensic features (the OCR and Loc. metrics). Removing CGAM also degrades all metrics to varying degrees.

Conclusion: CGAM is an important structure for improving forensic capability in joint forensic–calibration optimization, FQAC is the key design that further improves calibration, and the Stage 1 Pre-train process is the prerequisite for both designs to be effective. Every module and training stage makes its own contribution, so the full framework is necessary.

Table 1 (Q2): Stage-level ablation (average over four test sets)

ConfigurationPre-train (Stage1)SFT (Stage2)<conf> (Stage2)CGAM (Stage2)RforR_{\text{for}} (Stage3)RconfR_{\text{conf}} (Stage3)Cls.OCRLoc.Res.ECE↓
Pre-train✓–-–––38.8–40.9––
Pre-train + SFT✓✓-–––89.354.650.552.736.8
Pre-train + SFT + <conf>✓✓✓–––89.652.347.951.726.7
Pre-train + SFT + <conf> + CGAM✓✓✓✓––90.855.450.855.424.0
Pre-train + SFT + <conf> + CGAM + RforR_{\text{for}}✓✓✓✓✓–91.556.551.256.923.8
SFT + <conf> + CGAM + RforR_{\text{for}} + RconfR_{\text{conf}} (new)-✓✓✓✓✓70.619.618.926.039.7
Pre-train + SFT + <conf> + RforR_{\text{for}} + RconfR_{\text{conf}} (new)✓✓✓-✓✓90.255.851.054.622.3
Pre-train + SFT + <conf> + CGAM + RforR_{\text{for}} + RconfR_{\text{conf}}✓✓✓✓✓✓91.056.851.757.421.4
AuthorsOpenReview ↗

Response to Reviewer LxEU (Part 2)

Response to Q3: What Exactly Does Reasoning Confidence Capture, and How Does It Identify Unreliable Predictions?

We thank the reviewer for the suggestion.

  • EthinkE_{\text{think}} is an auxiliary proxy for the model's certainty in its own generated sequence. In Stage 2 (CGAM-guided aggregation) it is progressively optimized via hconfh_{\text{conf}} to capture the correctness of the model's forensic analysis process; in Stage 3 it is fused with AcrossA_{\text{cross}} and AcurA_{\text{cur}} to capture process and outcome correctness, respectively.
  • Based on the above analysis, reasoning confidence identifies unreliable predictions at two levels: (1) Stage 2 training enables it to distinguish samples whose analysis process is wrong; (2) Stage 3 uses the process-correctness information it carries as supervision to train the model, so that in an automated business workflow the model can proactively report low confidence for samples with uncertain analysis and route them to manual review to prevent misjudgments.

We have also added an empirical analysis, reported in Sec. 4.6 of the revised manuscript and in the appendix:

  • Experimental setup: In Table 2 (Q3), using Qwen3-VL-32B-Instruct with privileged ground-truth forensic knowledge (authenticity, tampered region/text/method, and traces) as a knowledge-informed reference, we correlate its teacher-forced per-token distribution with that of our 8B model over the same <think> sequence (Pearson ρ\rho; the shared tokenizer aligns positions directly). We then stratify samples and examine, with EthinkE_{\text{think}} rising, how ρ\rho and classification accuracy vary.

  • Result: ρ\rho rises monotonically across EthinkE_{\text{think}} quintiles (0.310 → 0.508, +0.198), while classification accuracy climbs from 88.2% to 98.0%. This indicates that EthinkE_{\text{think}} encodes whether the model's reasoning agrees with a correct, knowledge-informed forensic analysis.

Table 2 (Q3): ρ\rho and classification accuracy across EthinkE_{\text{think}} quintiles

Mean EthinkE_{\text{think}}Mean ρ\rho↑Cls. Acc (%, ↑)
0.62080.31088.24
0.62640.36490.20
0.67140.40288.24
0.71840.44491.18
0.72820.50898.04
Reviewer_zyUWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses an important issue. MLLMs can detect manipulated text in images; however, their predictions might not be reliable as they lack well-calibrated confidence estimates. To resolve this issue, the authors propose a three-stage framework, in which they first perform forensic continual pretraining, followed by SFT with CGAM, and finally GRPO reinforcement learning.

Critical strengths

1). The paper is well written, and the authors have put considerable effort into its readability and understandability.

2). The results seem promising: classification: 91%, OCR: 56.8%, localization: 51.7%, reasoning: 57.4%, and ECE: 21.4%.

Critical weaknesses

1). Computational complexity is not sufficiently discussed. 2). Difficulty in isolating improvements from the training pipeline. 3). Limited evidence of robustness to image degradation.

Questions for authors

1). The authors propose the target confidence as (w_1 E_{think} + w_2 A_{cur} + w_3 A_{cross}). I feel that a high token probability does not necessarily imply that the model has correctly interpreted the forensic evidence. Could the authors explain this further or provide some empirical evidence?

2). Given the reported ECE of 21.4%, could the authors further discuss the effectiveness of the proposed confidence mechanism for selective prediction?

3). If possible, I think it would be helpful if the authors could consider reporting results across multiple random seeds to better understand the consistency of the improvements in ECE and forensic performance (not necessary for the rebuttal).

4). Even though the authors provide an ablation study on the components of the target confidence in Table 5, I think it would be helpful to explain the choice of weights (0.50, 0.25, 0.25) in the main paper, particularly why reasoning certainty is assigned a higher weight than the other two components, or perhaps provide an additional ablation study.

5). I think it would be good to report the training cost and provide some inference latency analysis.

6). The authors have provided useful ablation studies to demonstrate the effectiveness of CGAM. I think an additional comparison with a simpler auxiliary classification head without the attention mechanism could further help clarify the contribution of the proposed architecture.

7). Even though the authors evaluate their method across multiple datasets, I feel that the robustness of the proposed approach under common image degradations, such as compression, blur, and noise, could be explored further.

4 comments
AuthorsOpenReview ↗

Response to Reviewer zyUW (Part 1)

Response to Q1: Does Reasoning Token Probability Reflect the Quality of Forensic Reasoning?

We thank the reviewer for this important question. We have added further analyses of what EthinkE_{\text{think}} captures to Sec 4.6 of the revised manuscript.

  • EthinkE_{\text{think}} is an auxiliary proxy for the model's certainty in its own generated sequence. In Stage 2 (CGAM-guided aggregation) it is progressively optimized via hconfh_{\text{conf}} to express the correctness of the model's forensic analysis process; in Stage 3 it is fused with AcrossA_{\text{cross}} and AcurA_{\text{cur}} to capture process and outcome correctness, respectively.

  • Experimental setup: In Table 1 (Q1), using Qwen3-VL-32B-Instruct with privileged ground-truth forensic knowledge (authenticity, tampered region/text/method, and traces) as a knowledge-informed reference, we correlate its teacher-forced per-token distribution with that of our 8B target over the same <think> sequence (Pearson ρ\rho; the shared tokenizer aligns positions directly). We then stratify samples and examine, with EthinkE_{\text{think}} rising, how ρ\rho and classification accuracy vary.

  • Result: ρ\rho rises monotonically across EthinkE_{\text{think}} quintiles (0.310 → 0.508, +0.198), while classification accuracy climbs from 88.2% to 98.0%. This indicates that EthinkE_{\text{think}} encodes whether the model's reasoning agrees with a correct, knowledge-informed forensic analysis.

Table 1 (Q1): ρ\rho and classification accuracy across EthinkE_{\text{think}} quintiles

Mean EthinkE_{\text{think}}Mean ρ\rho↑Cls. Acc (%, ↑)
0.62080.31088.24
0.62640.36490.20
0.67140.40288.24
0.71840.44491.18
0.72820.50898.04

Response to Q2: Can an ECE of 21.4% Support Selective Prediction?

We thank the reviewer for the suggestion. We have added a Selective Prediction experiment, reported in Sec. 4.3 of the revised manuscript.

  • Setup: test samples are ranked by the model's output confidence; a prediction is automatically accepted only when its confidence exceeds a threshold τ\tau, and the remaining samples are routed to manual review. We report the prediction risk Risk(τ)=1−Acc({i:ci≥τ})\mathrm{Risk}(\tau) = 1 - \mathrm{Acc}(\{i: c_i \geq \tau\}) at different Coverage levels (the fraction of automatically accepted samples);

  • As shown in Table 2 (Q2) below, compared with the Baseline and other calibration methods, our framework achieves the lowest prediction risk at every Coverage level.

  • In addition, we tested whether the full-set ECE of 21.4% carries over to the auto-accepted region by performing an ECE analysis of the auto-accepted region: at a fixed threshold τ=0.9\tau=0.9, only samples with C^≥0.9\hat{C}\geq 0.9 are retained as the auto-accepted set, and ECE is recomputed within this set using the same 10-bin binning scheme. The results show that this threshold automatically accepts 48.0% of the test samples, while the conditional ECE of the auto-accepted region is only 7.31%, far below the full-set value of 21.4%. In other words, the 21.4% ECE mainly falls in the low-confidence region excluded by the threshold and routed to manual review; it does not represent the calibration error of the samples actually auto-accepted in selective prediction.

Together, these two experiments show that the current ECE result is sufficient to support Selective Prediction in practical settings.

Table 2 (Q2): Prediction risk at different Coverage levels (Risk↓ = 1 − Accuracy; average over four test sets)

CoverageBaselineRLCRDCPOOurs
100%10.419.319.769.07
90%10.889.4310.049.25
80%11.639.8110.559.35
70%12.6310.0910.608.81
60%13.8310.7410.718.48
50%12.1810.3410.237.55

Response to Q3: Can Results Across Multiple Random Seeds Be Provided?

We thank the reviewer for the suggestion. We will add experimental results under multiple random seeds in the final revised version.

AuthorsOpenReview ↗

Response to Reviewer zyUW (Part 2)

Response to Q4: Why Are the FQAC Weights Set to (0.50, 0.25, 0.25)?

We thank the reviewer for the suggestion. First, we would like to correct a typo: the weights corresponding to EthinkE_{\text{think}}, AcurA_{\text{cur}}, and AcrossA_{\text{cross}} are 0.25, 0.25, and 0.50, respectively. We have corrected this error in the revised manuscript and added the motivation behind the weight design.

  • On the one hand, AcrossA_{\text{cross}}, as the average classification accuracy over multiple rollouts, expresses the stability of the model's predictions and is the key signal for assessing whether the model has sufficient discriminative ability on a given sample; by contrast, AcurA_{\text{cur}} may be subject to chance, so AcrossA_{\text{cross}} is assigned a larger weight.
  • On the other hand, since the ultimate goal of this work is to build a framework usable in automated business workflows, the model's output confidence should align as closely as possible with classification outcome accuracy, whereas EthinkE_{\text{think}} tends to express the correctness of the analysis process; it is therefore also inappropriate to give EthinkE_{\text{think}} a weight higher than that of AcrossA_{\text{cross}}.
  • We have added a weight-sensitivity experiment reporting both ECE and the four forensic metrics. By comparison, as shown in Table 3 (Q4) below, setting (w1,w2,w3)=(0.25,0.25,0.5)(w_1,w_2,w_3)=(0.25,0.25,0.5) achieves the best overall performance.

Table 3 (Q4): FQAC weight sensitivity (w1+w2+w3=1w_1+w_2+w_3=1; average over four test sets)

Settingw1w_1 (EthinkE_{\text{think}})w2w_2 (AcurA_{\text{cur}})w3w_3 (AcrossA_{\text{cross}})Cls.OCRLoc.Res.ECE↓
A0.000.500.5091.855.951.156.422.0
B0.3750.3750.2591.054.950.455.423.6
C0.250.250.5091.056.851.757.421.4
D0.1250.1250.7591.155.350.555.623.5
E1.000.000.0091.156.751.657.222.4

Response to Q5: What Are the Training Cost and Inference Latency?

We thank the reviewer for the suggestion. We have added a computational cost analysis, reported in the new Sec. 4.7 (Cost Analysis) of the revised manuscript.

  • Training cost: we report the number of epochs and the training time separately for each of the three stages in Table 4 (Q5) below. The epoch settings are already given in the original manuscript: Pre-train 1 epoch, SFT 2 epochs, GRPO 1 epoch (4 rollouts per prompt, temperature 1.0).
  • Inference latency: we report the average per-sample latency and the throughput in Table 5 (Q5) below.

Table 4 (Q5): Training cost

StageEpochsTraining Time
Pre-train115.2 h
SFT + CGAM23.4 h
GRPO + FQAC150.6 h
Total469.2 h

Table 5 (Q5): Inference latency (single GPU, batch=1)

MethodAvg. Latency (ms/sample)Throughput (samples/s)
Full Model59010.170

Response to Q6: Is CGAM Superior to a Simple Auxiliary Classification Head?

We thank the reviewer for the suggestion. We agree that the necessity of the Attention structure in CGAM requires further verification.

We have added new structural comparisons to Table 4 of the paper (four configurations in total), and report them in Table 6 (Q6) below: no auxiliary classification head (Baseline), classification after concatenating hconf/hts/hansh_{\text{conf}}/h_{\text{ts}}/h_{\text{ans}}, a learnable vector as the Query, and CGAM with hconfh_{\text{conf}} as the Query.

The results in Table 6 (Q6) show that, compared with direct concatenation followed by classification, the Attention structure yields gains on all four forensic metrics and both calibration metrics: Cls./OCR/Loc./Res. improve by +0.5/+2.4/+1.7/+3.3, while ECE/Brier decrease by −3.8/−2.2. This indicates that the Attention-based aggregation does lead to better optimization.

Table 6 (Q6): Comparison of CGAM structural variants (average over four test sets)

MethodQueryAggregationCls.OCRLoc.Res.ECE↓Brier↓
Baseline––89.652.347.951.726.718.3
Concatenation–hconf/hts/hansh_{\text{conf}}/h_{\text{ts}}/h_{\text{ans}}90.554.450.054.125.217.3
Learnable QueryLearnable vectorAttention90.654.950.756.122.115.2
CGAMhconfh_{\text{conf}}Attention91.056.851.757.421.415.1
AuthorsOpenReview ↗

Response to Reviewer zyUW (Part 3)

Response to Q7: Is the Method Robust Under Image Degradations?

We thank the reviewer for the suggestion.

We have added a degradation-robustness experiment in the appendix, applying three types of degradation uniformly across the four test sets:

  • JPEG compression with quality factor Q∈{90,70,50}Q \in \{90, 70, 50\};
  • Gaussian blur with σ∈{0.5,1.0,2.0}\sigma \in \{0.5, 1.0, 2.0\};
  • Gaussian noise with σ∈{5,10,20}\sigma \in \{5, 10, 20\} in pixel space [0,255][0,255].

Table 7 (Q7) below reports the four forensic metrics and ECE for each degradation type and severity level.

Table 7 (Q7): Robustness to image degradation (average over four test sets)

DegradationSeverityCls.OCRLoc.Res.ECE↓
Clean091.056.851.757.421.4
JPEG9091.056.651.156.722.5
JPEG7087.451.146.652.324.1
JPEG5086.649.445.150.525.5
Gaussian Blur0.590.657.051.456.222.2
Gaussian Blur1.077.052.449.145.622.3
Gaussian Blur2.068.444.541.731.227.7
Gaussian Noise584.755.550.855.419.0
Gaussian Noise1078.651.748.050.419.7
Gaussian Noise2075.846.543.047.421.0

I thank the authors for the detailed rebuttal. It addressed my major concerns, and I'll keep my score as it is.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank the reviewers for their constructive feedback, which helped us clarify both the methodological details and the presentation of our experimental evidence. The revisions fall into three areas:

1. Presentation

  • Sec. 3.4 (Reinforcement Learning), Eq. (6) — We reordered the weights in the target-confidence equation to C=w1Across+w2Acur+w3EthinkC=w_1A_{\text{cross}}+w_2A_{\text{cur}}+w_3E_{\text{think}}, so that symbol w1w_1 corresponds to the cross-rollout correctness term that receives the largest weight.
  • Sec. 3.4 (Reinforcement Learning), Eq. (6) — We added an explanatory sentence after the equation stating that AcrossA_{\text{cross}} captures the stability of verdict correctness and therefore receives the largest weight, and clarified what each of w1,w2,w3w_1,w_2,w_3 controls.
  • Sec. 3.4 (Reinforcement Learning), Eq. (8) — We added a note below Eq. (8) on the implementation detail of the Format reward, which uses a regular expression to check whether the output strictly matches <think>…</think><answer>…</answer><conf>…</conf> and returns 1 if valid and 0 otherwise.

2. Experimental Updates and Additional Analyses

  • Sec. 4.3 (Calibration Comparison), Table 2 — We restructured the calibration table into averaged results over ECE, Brier score, and AUROC (replacing the previous per-test-set ECE table), together with the Δ\Delta relative to the baseline.
  • Sec. 4.3 (Calibration Comparison) — We added a selective-prediction analysis: with the auto-accepted set defined as C^≥0.9\hat C\ge 0.9 (covering 48.0%48.0\% of samples), ECE within this subset is only 7.3%7.3\%.
  • Sec. 4.3 (Calibration Comparison), Table 3 — We added a risk–coverage table comparing the prediction risk of automatically accepted samples at coverage levels from 100%100\% to 50%50\%.
  • Sec. 4.4 (Ablation Study), Table 4 — We added two rows ("without pretraining" and "without CGAM") in the component-ablation table to isolate these two factors.
  • Sec. 4.4 (Ablation Study) — We extended the accompanying analysis of the component ablation accordingly, showing that Stage 1 is a prerequisite for joint optimization.
  • Sec. 4.4 (Ablation Study), Table 5 — We expanded the confidence-query ablation table from 2 to 4 rows (Baseline, Concat., Learnable, hconfh_{\text{conf}}), showing that hconfh_{\text{conf}} is best on all metrics.
  • Sec. 4.6 (What Reasoning Certainty Captures), Table 7 — We added a new main-text subsection, "What Reasoning Certainty Captures," which uses the Pearson correlation with a knowledge-informed Qwen3-VL-32B reference model to characterize what EthinkE_{\text{think}} actually tracks.

3. Appendix Updates

  • Sec. A.2.1 (Robustness to Image Degradations), Table 9 — We added a new "Robustness to Image Degradations" subsection with a corresponding table covering forensic metrics and ECE under JPEG compression, Gaussian blur, and Gaussian noise.
  • Sec. A.2.4 (FQAC Weight Sensitivity), Table 11 — We added a new "FQAC Weight Sensitivity" subsection with a corresponding table comparing five weight configurations under w1+w2+w3=1w_1+w_2+w_3=1, confirming that Setting C is optimal.
  • Sec. A.2.5 (Computational Cost Analysis), Tables 12–13 — We added a new "Computational Cost Analysis" subsection reporting the per-stage training time (69.1 h in total) and single-GPU inference latency (5901 ms).
  • Sec. A.2 (More Experiment) — We reordered the appendix subsubsections, moving the degradation-robustness experiment earlier so that the experimental content is grouped more coherently.

We again thank the reviewers for their constructive comments; the additional analyses and clarifications above were directly motivated by their feedback, and we hope the revised version presents the motivation and evidence of our method more clearly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.