acceptodds
Under review as a conference paper at ICLR 2027

Visual Thought Matters: Learning to Ground and Refine Continuous Reasoning for Generalizable Deepfake Detection

Abstract

Generalizable deepfake detection requires reasoning over subtle visual evidence that is difficult to faithfully express in discrete language. While multimodal large language models (MLLMs) can generate detailed reasoning chains, text-dominated reasoning may overlook fine-grained forensic cues. To this end, we propose COVE, a framework that integrates continuous forensic states into the reasoning trajectory to capture and leverage fine-grained visual evidence throughout reasoning. Complementary frozen visual experts supervise these states to encode appearance, semantic, and boundary cues, equipping the MLLM with fine-grained forensic perception. To encourage the model to use this evidence for decision making, we introduce a Counterfactual Token Utility Reward (CTUR), which quantifies the contribution of forensic states to the correct verdict through counterfactual suppression and optimizes their decision utility via reinforcement learning. We further introduce a contrastive grounding objective that contrasts the same target reasoning under semantically matched visual conditions and optimizes its grounding gain relative to a frozen reference model. This encourages the refined reasoning to depend on its supporting visual evidence, while complete-chain supervision and replay preserve reliable reasoning behavior. At inference, COVE uses a single MLLM to generate the complete forensic reasoning trajectory and authenticity verdict. Experiments on HydraFake demonstrate strong generalization across diverse out-of-distribution scenarios, particularly on unseen forgeries and data domains, while producing high-quality forensic reasoning. Together, these results demonstrate the effectiveness of COVE in connecting fine-grained visual representation, decision utility, and evidence-grounded reasoning for generalizable deepfake detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Generalizable deepfake detection, Multimodal large language models, Continuous latent reasoning

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_Bk6HOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes COVE, a framework for improving deepfake detection by combining visual evidence with multimodal language model reasoning. It introduces continuous visual states supervised by three pretrained visual models to capture subtle image details. A counterfactual reward encourages the model to use these visual states when making decisions, while contrastive training aims to keep its reasoning grounded in the image.

Critical strengths
  1. Using different models to supervise continuous visual states is a reasonable way to combine different types of visual information, including appearance, semantics, and boundaries.

  2. The proposed counterfactual reward is interesting because it goes beyond learning visual features and explicitly encourages the model to use them when making authenticity decisions.

  3. The experiments cover several types of distribution shifts, and the ablation studies provide useful evidence that the visual experts, counterfactual reward, and reasoning refinement contribute to the improvements.

Critical weaknesses
  1. The counterfactual reward replaces forensic token embeddings with an average vocabulary embedding, which may introduce unnatural inputs rather than isolate the contribution of meaningful visual evidence.

  2. The contrastive training pairs are created using generated images, but it is unclear whether they preserve enough visual details to ensure that authenticity is the main difference between the two images.

  3. The paper mainly evaluates reasoning quality using other language models as judges, without directly verifying whether the visual evidence described in the explanations is actually present in the images.

Questions for authors
  1. How can the authors verify that the counterfactual reward measures the actual contribution of forensic evidence, rather than the effects of replacing tokens with unnatural embeddings?

  2. Can the authors provide a more direct evaluation of whether the generated explanations are supported by real visual evidence, rather than simply being plausible explanations associated with correct predictions?

8 comments
AuthorsOpenReview ↗

Response to Reviewer Bk6H

Dear Reviewer Bk6H,

We sincerely thank you for the time and effort you have devoted to reviewing our paper. Your constructive feedback is invaluable in helping us improve the clarity and rigor of our work. Below, we provide point-by-point responses to your comments. We have also incorporated all corresponding improvements into the revised manuscript, with all changes highlighted in blue for ease of reference. Please feel free to let us know if you have any further concerns or questions.


W1 / Q1: CTUR with alternative replacement strategies

We thank the reviewer for pointing out that the average vocabulary embedding may differ in distribution from the original forensic tokens. To examine whether the training gains from CTUR depend on this particular replacement strategy, we add two counterfactual interventions: sampling from the embedding distribution of each expert’s tokens, and replacing forensic states with those from trajectories generated for other images. These experiments examine the sensitivity of CTUR to the source of the replacement vectors. We use each replacement strategy to compute CTUR in a separate GRPO training run, and evaluate the resulting models on normal inputs without intervention.

Table 1. CTUR training results with different counterfactual interventions (%). Exp1: sampling from expert-specific token embedding distributions; Exp2: substituting states from another image; Exp3: average vocabulary embedding (Ours).

Replacement strategy in CTURIDCMCFCDOverall
w/o CTUR97.198.793.280.791.2
Exp197.199.096.383.193.0
Exp297.299.396.283.093.1
Exp397.299.196.383.293.1

As shown in Table 1, the three interventions achieve similar detection performance, with overall accuracy ranging from 93.0% to 93.1%. This represents an improvement of 1.8%–1.9% over the 91.2% achieved without CTUR. These results show that the training gains from CTUR are observed across different replacement strategies and are not limited to the average vocabulary embedding.

Given their similar performance, we use the average vocabulary embedding as the default implementation because it can be computed in advance and reused directly. It requires neither additional distribution sampling nor trajectory generation for other images, making the replacement procedure simpler. We have included these comparisons in the Appendix H.


W2: Semantic consistency of paired images

We thank the reviewer for raising this concern. For the contrasting image conditions, we use pairs of real source images and their corresponding fake images from the HydraFake training set. We pair each fake image with its real source image using the existing source records. This provides a traceable basis for their content correspondence while reducing unrelated visual differences. The refinement stage uses only the existing HydraFake training data and does not generate additional data. The content correspondence between the paired training samples follows from their generation or manipulation relationship, which helps reduce the influence of unrelated differences in subjects and scenes on the objective. We have corrected the descriptions of generating new images in Section 3.4 and Appendix A.3.

W3 / Q2: Evaluation of visual evidence in model explanations

We thank the reviewer for the suggestion. For a fair comparison, we follow the evaluation protocols of Veritas[1] and AIGI-Holmes[2], both of which emphasize interpretability and the grounding of visual evidence in the input image. The judge model receives the image, the ground-truth label, and the complete response, and is explicitly instructed to consider whether the evidence corresponds to the image. Thus, the existing evaluation includes visual grounding in its scoring criteria. The prompt template is provided in Appendix D of the revised manuscript.

To further assess the factual correctness of the explanations, we add a comparison using an additional external judge and a newly designed prompt template, also provided in Appendix D. This prompt assigns separate scores from 1 to 5 for Visual Grounding, Evidence Specificity, and Reasoning Soundness. We average the three scores to obtain the final score for each sample, placing greater emphasis on whether the model’s reasoning contains actual visual evidence.

Table 2. Evaluation of reasoning quality based on visual evidence.

ModelScore Evaluation (GPT-6-Sol)
MiMo-VL-7B2.901
GPT-4o2.665
Gemini-2.5-Pro3.134
Veritas4.015
COVE (cold-start)3.489
COVE (base stage)4.041
COVE (ours)4.151

As shown in Table 2, COVE continues to achieve a high score with the new prompt template and the more advanced external judge, GPT-6-Sol, indicating that its reasoning is supported by actual evidence from the corresponding image for most samples. Appendix D of the revised manuscript provides an example of the external judge’s output.


[1] Veritas: Generalizable Deepfake Detection via Pattern-Aware. ICLR 2026.

[2] AIGI-Holmes: Towards Explainable and Generalizable AI-Generated Image Detection via Multimodal Large Language Models. ICCV 2025.

Dear Reviewer, The authors have provided their responses to the raised concerns and revisions. Please review their responses and join the discussion before the end of the author-review discussion phase (Nov 18, AoE). Thanks.

Best, AC

Reviewer_PRLnOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes COVE, a framework for generalizable deepfake detection that integrates continuous forensic states into multimodal reasoning. The method consists of three main components: (1) learning continuous visual tokens supervised by complementary frozen experts (DINOv3, SigLIP2, and PiDiNet), (2) introducing a Counterfactual Token Utility Reward (CTUR) to encourage the model to utilize these representations for authenticity prediction, and (3) refining the reasoning process through contrastive forensic grounding with correction and replay.

Experiments on HydraFake demonstrate strong generalization across unseen generators, forgery types, and image domains, achieving 94.0% overall accuracy. Additional evaluations on broader benchmarks, visual corruptions, and reasoning quality further support the effectiveness of the framework.

Critical strengths
  1. A well-motivated and coherent framework for evidence-aware forensic reasoning.

The paper addresses an important distinction between encoding visual information and actually utilizing it during reasoning. Although continuous visual tokens and expert-guided representation learning have been explored in prior work such as CoVT, the proposed CTUR provides an explicit optimization signal for encouraging the use of forensic states in authenticity decisions. The subsequent contrastive grounding objective further attempts to connect reasoning refinement with visual evidence. These components form a coherent framework with clear motivations.

  1. Strong empirical performance and comprehensive experimental validation.

COVE achieves competitive or superior results against recent MLLM-based forensic methods, particularly under cross-forgery and cross-domain evaluation. The ablation studies provide useful evidence for the contributions of the visual experts, CTUR, and reasoning refinement. The evaluation also covers broader image forgery benchmarks, compression and blur robustness, and reasoning quality, making the empirical assessment relatively comprehensive.

Critical weaknesses
  1. The methodological novelty is somewhat incremental relative to existing visual reasoning approaches.

The continuous visual token design and expert supervision are closely related to CoVT, while the counterfactual and reference-relative optimization objectives build upon established ideas in reinforcement learning and visual grounding. The main contribution therefore lies in adapting and integrating these techniques for forensic reasoning rather than introducing an entirely new reasoning architecture. Nevertheless, the task-specific motivation and empirical effectiveness make this integration meaningful.

  1. The evidence for improved visual grounding could be more direct.

CTUR measures how suppressing forensic tokens changes the likelihood of the correct verdict, but this dependency does not necessarily establish that the model utilizes semantically meaningful forensic evidence. Similarly, the contrastive grounding objective encourages image-dependent reasoning, but its effectiveness in improving the factual correctness of individual forensic explanations is primarily supported by indirect metrics and qualitative examples. More targeted analyses would strengthen the central claims.

Questions for authors

Regarding CTUR: Could the authors further clarify whether the observed counterfactual performance degradation primarily reflects the removal of meaningful forensic information rather than distribution shifts introduced by replacing dedicated visual token embeddings? A comparison with distribution-matched or unrelated visual token replacements would help establish the interpretation of CTUR.

Regarding Contrastive Forensic Grounding: How are the matched and contrasting image conditions defined for correction examples with different ground-truth authenticity labels? In particular, how does the implementation ensure that the preferred visual condition consistently supports the target reasoning? Additionally, how is semantic consistency between the original and generated images controlled to prevent unrelated visual differences from influencing the objective?

2 comments
AuthorsOpenReview ↗

Response to Reviewer PRLn

Dear Reviewer PRLn,

Thank you for your insightful comments and positive feedback, which are highly encouraging to us. We sincerely appreciate your valuable suggestions and questions. Below, we provide point-by-point responses to your concerns, and all corresponding revisions have been incorporated into the manuscript and highlighted in blue for ease of reference. If you have any further questions, please do not hesitate to contact us.


Q1: CTUR with alternative replacement strategies

We thank the reviewer for raising the concern about possible distribution shifts. To examine whether the training gains from CTUR depend on this particular replacement strategy, we add two counterfactual interventions: sampling from the embedding distribution of each expert’s tokens, and replacing forensic states with those from trajectories generated for other images. These experiments examine the sensitivity of CTUR to the source of the replacement vectors. We use each replacement strategy to compute CTUR in a separate GRPO training run, and evaluate the resulting models on normal inputs without intervention.

Table 1. CTUR training results with different counterfactual interventions (%). Exp1: sampling from expert-specific token embedding distributions; Exp2: substituting states from another image; Exp3: average vocabulary embedding (Ours).

Replacement strategy in CTURIDCMCFCDOverall
w/o CTUR97.198.793.280.791.2
Exp197.199.096.383.193.0
Exp297.299.396.283.093.1
Exp397.299.196.383.293.1

As shown in Table 1, the three interventions achieve similar detection performance, with overall accuracy ranging from 93.0% to 93.1%. This represents an improvement of 1.8%–1.9% over the 91.2% achieved without CTUR. These results show that the training gains from CTUR are observed across different replacement strategies and are not limited to the average vocabulary embedding.

Given their similar performance, we use the average vocabulary embedding as the default implementation because it can be computed in advance and reused directly. It requires neither additional distribution sampling nor trajectory generation for other images, making the replacement procedure simpler. We have included these comparisons in the Appendix H.


Q2: Matched image conditions and semantic consistency

We thank the reviewer for pointing out this issue in our description. x+x^+ denotes the image matched to the target teacher chain and does not necessarily denote a real image. For a fake sample, the original fake image provides the matched image condition, and its real source image provides the contrasting image condition. We have corrected the original description that defined x+x^+ as a real image in all cases.

To address semantic consistency, we use pairs of real source images and their corresponding fake images from the HydraFake training set. We pair each fake image with its real source image using the existing source records. This provides a traceable basis for their content correspondence while reducing unrelated visual differences. The refinement stage uses only the existing HydraFake training data and does not generate additional data. The content correspondence between the paired training samples follows from their generation or manipulation relationship, which helps reduce the influence of unrelated differences in subjects and scenes on the objective.

Thank you for the detailed response and additional experiments. The new results and clarifications substantially address my concerns. I appreciate the authors' efforts and will maintain my original score.

Reviewer_dkAVOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

COVE extends the Veritas detector (InternVL3-8B) with three additions. First, twelve forensic tokens are inserted into the reasoning trace and aligned to frozen DINOv3, SigLIP2 and PiDiNet experts. Second, CTUR is a GRPO reward that measures how much the correct-label likelihood drops when those token embeddings are suppressed. Third, a refinement stage contrasts the same teacher chain under a real image and a generated counterpart. COVE reports 94.0% overall accuracy on HydraFake, compared with 92.1% for VIGIL and 90.7% for Veritas.

Critical strengths
  1. The paper separates three questions: what the evidence encodes, whether the verdict uses it, and whether the reasoning depends on the image. This is a useful decomposition.

  1. CTUR is a novel reward that targets the interventional effect of internal states on the verdict. Regenerating the continuation after the intervention captures downstream influence.

  1. COVE beats Veritas, which uses the same backbone, by +3.3 overall and +6.8 on cross-forgery, and it also beats VIGIL. The CTUR ablation is properly controlled from the same initialization.

  1. Experts are used only in training, so inference adds no expert cost. Per-subset precision and recall, correction and regression counts, and failure cases are reported.
Critical weaknesses
  1. The CTUR gain may not come from the counterfactual at all. Its first term, log P(a*|c^F), rewards confidence in the correct label, which is a strong training signal on its own. The paper never trains with this term alone, so the gain cannot be attributed to the counterfactual.

  1. The CTUR intervention does not remove visual evidence. Only the fixed input embedding of each special token is replaced, and the hidden states still attend to the image. CTUR therefore seems to measures sensitivity to an unusual input, not reliance on image content.

  1. The expert ablation overstates the experts' contribution. Its "no expert" row exactly matches Veritas's published numbers, and its full row also includes CTUR. In a fair comparison without CTUR, the experts add only +0.5 overall and hurt cross-domain accuracy (80.7 vs 82.2).

  1. The refinement results are weak evidence. The stage adds only +0.9 overall from a single seed, and half of that comes from DeepFaceLab. All hyperparameters are tuned on the test sets, and the paper does not say which model generated x⁻ or the teacher chains, so overlap with test generators cannot be ruled out.

  1. The reasoning is often not grounded in the image, contrary to the paper's claim. The qualitative examples call a polar bear "feline", analyze faces on a document page, describe a side-lit face as evenly lit, and reuse the same text across images. In Figure 5 (last image example), the reasoning seems completely hallucinated. The LLM judge sees the label and never checks whether the stated evidence is true.
Questions for authors
  1. Are the forensic positions discrete special tokens with fixed input embeddings at inference? If so, why call this continuous reasoning?

  1. What is test-time accuracy when forensic states are mean-replaced, or swapped from another image?

  1. Which generator produces x⁻, and does it overlap with any HydraFake test generator? Do the teacher chains see the label?

  1. Why do linear probes on DINOv3 and SigLIP2 features reach only 55%, while UniFD reaches 78% on the same benchmark?

  1. Why does COVE (cold-start) receive a higher Elo than Veritas while scoring lower with both judges?

  1. Can you compare against AIGI-Holmes, FakeReasoning, X2-DFD and a CoVT-style baseline, and report training compute versus Veritas?

Please check the weakness section too.

9 comments
AuthorsOpenReview ↗

Response to Reviewer dkAV

Dear Reviewer dkAV:

We sincerely thank you for the time and effort you have devoted to reviewing our paper. Your constructive feedback is invaluable in helping us improve the clarity and rigor of our work. Below, we provide point-by-point responses to your comments. We have also incorporated all corresponding improvements into the revised manuscript, with all changes highlighted in blue for ease of reference. Please feel free to let us know if you have any further concerns or questions.


W1: Contribution of the counterfactual term

We thank the reviewer for pointing out the need to distinguish the contributions of the reward for the correct label and the counterfactual comparison term. To directly examine this issue, we add an ablation that retains only log⁡Pθ(a∗∣x,u,cF)\log P_\theta(a^*\mid x,u,c^F) as the reward, and compare it with the settings without CTUR and with full CTUR.

All three experiments start from the same DPO checkpoint and use the same training data, GRPO configuration, other rewards, and evaluation settings. Evaluation is conducted after GRPO training. In the additional control experiment, we replace full CTUR with the log-probability of the correct label as the GRPO reward signal, keeping its original reward weight.

Table 1. Detection accuracy (%) for the ablation of CTUR reward components.

RL reward settingIDCMCFCDOverall
RL w/o CTUR97.198.793.280.791.2
RL with log-probability reward97.198.993.281.091.4
RL with full CTUR (COVE base stage)97.299.196.383.293.1

As shown in Table 1, using the reward for the correct label increases Overall accuracy from 91.2% to 91.4%, a gain of 0.2%. In comparison, full CTUR achieves 93.1% Overall accuracy, exceeding this control by 1.7%. This comparison shows that the reward for the correct label does not achieve the full benefit of CTUR, and that the counterfactual comparison term provides additional benefits under the current training setup.

This result is consistent with the motivation behind CTUR. The reward log⁡Pθ(a∗∣x,u,cF)\log P_\theta(a^*\mid x,u,c^F) only measures how strongly the current reasoning context supports the correct label. For two trajectories with similar probabilities of the correct label, this reward cannot further distinguish their reliance on the forensic states. Full CTUR compares the change before and after intervention. Therefore, even trajectories with similar probabilities of the correct label can receive different rewards depending on the changes after intervention. This provides an additional signal for optimizing the contribution of the designated forensic positions to subsequent reasoning. We have reported this in Appendix H.

W2: CTUR with alternative replacement strategies

We thank the reviewer for pointing out that the average vocabulary embedding may differ in distribution from the original forensic token representations. To examine whether the training gains from CTUR depend on this particular replacement strategy, we add two counterfactual interventions: sampling from the embedding distribution of each expert’s tokens, and replacing forensic states with those from trajectories generated for other images. These experiments examine the sensitivity of CTUR to the source of the replacement vectors. We use each replacement strategy to compute CTUR in a separate GRPO training run, and evaluate the resulting models on normal inputs without intervention.

Table 2. CTUR training results with different counterfactual interventions (%). Exp1: sampling from expert-specific token embedding distributions; Exp2: substituting states from another image; Exp3: average vocabulary embedding (Ours).

Replacement strategy in CTURIDCMCFCDOverall
w/o CTUR97.198.793.280.791.2
Exp197.199.096.383.193.0
Exp297.299.396.283.093.1
Exp397.299.196.383.293.1

As shown in Table 2, the three interventions achieve similar detection performance, with overall accuracy ranging from 93.0% to 93.1%. This represents an improvement of 1.8%–1.9% over the 91.2% achieved without CTUR. These results show that the training gains from CTUR are observed across different replacement strategies and are not limited to the average vocabulary embedding.

Given their similar performance, we use the average vocabulary embedding as the default implementation because it can be computed in advance and reused directly. It requires neither additional distribution sampling nor trajectory generation for other images, making the replacement procedure simpler. We have included these comparisons in the Appendix H.


W3: Contributions of expert supervision and CTUR

We thank the reviewer for raising the issue. We first clarify the configuration without experts: it removes the 12 added forensic tokens and their corresponding expert supervision while retaining the model’s original image input and reasoning framework. Since the reasoning trajectory no longer contains the designated forensic positions on which to intervene, the GRPO stage also excludes CTUR. This configuration therefore corresponds to the base Veritas setting, while the full configuration introduces both expert supervision and CTUR. The performance difference between them reflects the joint contribution of these two components.

Without CTUR, the configuration with experts improves Overall accuracy by 0.5% and CF accuracy by 2.9% over Veritas, but reduces CD accuracy by 1.5%. This is consistent with the motivation behind CTUR. Expert supervision trains the model to produce designated forensic states that encode visual information, while an additional reward signal is needed to encourage subsequent reasoning to make effective use of these states.

W4

W4.1: Refinement gains

We thank the reviewer for this comment. In addition to DeepFaceLab, the refinement stage improves accuracy on the CF subsets StarGANv2 and InfiniteYou from 97.3% to 98.4% and 98.7%, respectively, corresponding to relative reductions in error rate of approximately 40.7% and 51.9%. These two subsets involve face editing and identity-preserving personalized image generation unseen during training, respectively. The former requires identifying subtle editing artifacts in low-resolution portraits, while the latter preserves a natural facial appearance and coherent scene semantics, placing greater demands on the identification of fine-grained visual evidence. On these two subsets, the final COVE model outperforms VIGIL, the strongest existing baseline, by 2.9% and 3.8%, respectively. These results show that refinement can further improve the detection of unseen forgery types even from already high accuracy.

W4.2: Hyperparameter selection

Hyperparameters were selected on the validation set, and the test sets were used for evaluation after fixing the configuration. The original wording inaccurately linked test performance to configuration selection, and we have corrected this description in the Appendix F.

W4.3 / Q3: Sources of image pairs and teacher chains

For the contrasting image conditions, we use pairs of real source images and their corresponding fake images from the HydraFake training set. We pair each fake image with its real source image using the existing source records, and both images belong to the training set. This provides a traceable basis for their content correspondence while reducing unrelated visual differences. The refinement stage uses only the existing HydraFake training data and does not generate additional data. Therefore, the training generators do not overlap with those in the test sets other than ID. We have corrected the descriptions of generating new images in Section 3.4 and Appendix A.3.

The teacher chains are the complete reasoning annotations and real/fake answers already included in the original SFT training data. No additional teacher chains were generated. The original HydraFake annotation process includes the ground-truth authenticity labels of the training samples.


W5: Visual grounding of model explanations

We thank the reviewer for pointing this out. In some of the examples presented, the model predicted the correct answer, but the visual evidence described in its reasoning was not fully accurate. We have updated the examples to show cases with correct answers and reasoning grounded in the image. See Figures 4–7 in the revised manuscript.

For a fair comparison, we follow the evaluation protocols of Veritas[1] and AIGI-Holmes[2], both of which emphasize interpretability and the grounding of visual evidence in the input image. The judge model receives the image, the ground-truth label, and the complete response, and is explicitly instructed to consider whether the evidence corresponds to the image. Thus, the existing evaluation includes visual grounding in its scoring criteria. The prompt template is provided in Appendix D of the revised manuscript.

To further assess the factual correctness of the explanations, we add a comparison using an additional external judge and a newly designed prompt template, also provided in Appendix D. This prompt assigns separate scores from 1 to 5 for Visual Grounding, Evidence Specificity, and Reasoning Soundness. We average the three scores to obtain the final score for each sample, placing greater emphasis on whether the model’s reasoning contains actual visual evidence.

Table 3. Evaluation of reasoning quality based on visual evidence.

ModelScore Evaluation (GPT-6-Sol)
MiMo-VL-7B2.901
GPT-4o2.665
Gemini-2.5-Pro3.134
Veritas4.015
COVE (cold-start)3.489
COVE (base stage)4.041
COVE (ours)4.151

As shown in Table 3, COVE continues to achieve a high score with the new prompt template and the more advanced external judge, GPT-6-Sol, indicating that its reasoning is supported by actual evidence from the corresponding image for most samples. Appendix D of the revised manuscript provides an example of the external judge’s output.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.