acceptodds
Under review as a conference paper at ICLR 2027

CF-GRPO: Counterfactual Group Relative Policy Optimization for Video Hallucination Suppression

Abstract

Video language models have made substantial progress in video understanding, yet they remain prone to hallucination, producing answers that are inconsistent with the video content. Reinforcement fine-tuning with verifiable outcome rewards has been explored to mitigate such hallucinations. However, these rewards cannot distinguish visually grounded correct answers from those produced by the language prior. We show that this source-blind credit assignment is inherent in outcome-reward GRPO: the expected update is determined by the probability of correctness, regardless of whether that correctness depends on visual evidence. Our analysis further motivates interventions at three complementary levels of the GRPO update dynamics: the direction of pair-level updates, the magnitude of sample-level reinforcement, and the allocation of credit within each output group. Accordingly, we propose CF-GRPO, with three mechanisms targeting these respective levels. Specifically, 1)At the pair level, counterfactual pair optimization jointly trains opposite-label prompts sharing the same video, cancelling the common answer-shift update of prompts that the policy answers alike while preserving a positive discriminative update. 2)At the sample level, blind-failure advantage shaping scales the GRPO advantage by the model's failure probability without the video, reducing reinforcement for samples already solvable from the language prior. 3)At the within-group level, an evidence-consistency reward distinguishes outputs with identical correctness by whether their reasoning states answer-consistent evidence, restoring learning signal when outcome-reward GRPO provides none. Experiments on three video hallucination benchmarks show that CF-GRPO consistently improves over GRPO and strong baselines, while preserving the model’s answer distribution and general video understanding.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Video Hallucination, Multimodal large language models, Reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7nTEOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies video hallucination in video-language models and identifies a limitation of standard outcome-reward GRPO, termed source-blind credit assignment: a correct answer receives the same reinforcement whether it is grounded in the video or can already be produced from the model’s language prior. To address this, the authors propose CF-GRPO, which modifies the GRPO update at three complementary levels. At the pair level, counterfactual pair optimization jointly trains opposite-label prompts sharing the same video so that common answer-shift updates cancel while discriminative updates are preserved. At the sample level, blind-failure advantage shaping estimates how likely the model is to answer correctly without visual input and downweights samples that are already solvable from language priors. At the within-group level, an evidence-consistency reward reallocates credit according to whether the reasoning contains answer-consistent visual evidence. Experiments on Qwen2.5-VL-7B across EventHallusion, VideoHallucer, and VidHalluc show that CF-GRPO improves over standard GRPO and other baselines while largely preserving general video-understanding performance, with additional analyses indicating reduced reliance on answers that remain unchanged when the video is removed.

Critical strengths
  1. The paper identifies an important failure mode of outcome-based GRPO and provides a clear way to quantify visual dependence. The distinction between a correct answer that is genuinely grounded in the video and one that can be produced from the language prior is highly relevant for video hallucination. I particularly find the blind-input formulation useful: estimating whether the model can answer correctly without the video provides a direct and interpretable signal for how much a sample actually depends on visual evidence. This makes the source-blind credit assignment problem concrete than purely conceptual.

  2. The method is well structured and experimentally well supported. The three components of CF-GRPO target complementary parts of the optimization process—pair-level update direction, sample-level reinforcement magnitude, and within-group credit allocation—and the Evidence-Consistency Reward is especially useful for distinguishing outputs that have the same final correctness but differ in whether their reasoning is consistent with the visual evidence. The paper also provides strong empirical validation through matched GRPO comparisons, ablations, answer-distribution analysis, zero-video dependence analysis, and general video-understanding benchmarks, making the overall engineering contribution convincing.

Critical weaknesses
  1. It remains unclear how much of the improvement is specific to GRPO or to the counterfactual training recipe itself. CF-GRPO is compared carefully against standard GRPO, but the paper does not include a matched non-RL baseline trained with the same counterfactual pairs, hard negatives, evidence supervision, and blind-failure signal. In particular, the blind-failure weight 1−b(q)1-b(q) can also be interpreted as an adaptive curriculum or sample-weighting mechanism, and the counterfactual pairs themselves provide strong discriminative supervision. The current SFT baseline is only the initialization from which the RL variants are trained, not a counterfactual/curriculum SFT baseline using the same data. This makes it difficult to determine whether the gains arise from GRPO-specific credit assignment or from the underlying data construction and curriculum. A matched comparison against counterfactual SFT or weighted/curriculum SFT would substantially strengthen the methodological claim.

  2. The conceptual novelty is somewhat limited, with the contribution appearing primarily as a strong integration of existing ideas into GRPO. The observation that multimodal models can produce correct answers from language priors without relying on visual input is already well established, and counterfactual training has also been explored extensively for hallucination mitigation. The paper’s novelty therefore lies more in organizing these ideas into a unified GRPO framework—targeting update direction, sample-level reinforcement magnitude, and within-group credit allocation—than in introducing a fundamentally new learning principle. I view this as a well-executed engineering and methodological contribution with strong experiments, but the lack of matched non-GRPO baselines makes the incremental conceptual novelty more consequential to my overall assessment.

Questions for authors

*Could the authors compare CF-GRPO against other curriculum-learning such as VideoHallu? VideoHallu provides a carefully constructed expert dataset and benchmark, together with a curriculum-based training strategy for improving video understanding and reducing hallucination. Since several key ingredients of CF-GRPO—such as counterfactual samples and emphasizing examples that require stronger visual grounding—could also be incorporated into an curriculum-learning framework. A comparison using the same initialization, training data, and comparable training budget would be particularly informative. If CF-GRPO can consistently outperform such a strong curriculum-learning SFT/RL baseline, it would provide much stronger evidence for the specific advantage of Counter Factual setups, and I am more than happy to raise my score accordingly.

3 comments

We thank the reviewer for the detailed and constructive comments and for acknowledging that our work identifies an important failure mode of outcome-reward GRPO and that the method is well structured and experimentally well supported. We respond to each point below and include the requested matched baselines.

W1. Can matched comparisons against counterfactual SFT, weighted SFT and a curriculum-learning RL baseline such as VideoHallu be added, with the same initialization and training data?

Yes. As shown in Table R1, CF-GRPO outperforms matched RFT, weighted RFT, and curriculum GRPO on the overall metric of all three benchmarks. All baselines use the same SFT initialization and CF-GRPO training prompts. The baselines are implemented as follows.

  1. Non-RL baselines. RFT fine-tunes on correct model-generated responses, prioritizing those mentioning the evidence key. Weighted RFT additionally scales the loss by the prompt’s blind-failure probability, incorporating the same sample-weighting signal as CF-GRPO.
  2. Curriculum RL. Following the curriculum-based training strategy of VideoHallu, curriculum GRPO uses the same number of training steps as CF-GRPO and presents the base samples first and the counterfactual pairs afterward.

CF-GRPO exceeds the strongest of these baselines by 1.95, 1.57, and 0.72 percentage points on EventHallusion, VideoHallucer, and VidHalluc, respectively, supporting its benefit beyond the shared data and the tested weighting and curriculum strategies.

EventHallusionVideoHallucerVidHalluc
MethodEntireMixMisl.OverallObj.Sem.Ext.OverallAccYesNoACH
SFT (initialization)47.3749.2275.4955.2663.5064.5057.0052.6086.0689.5282.6073.50
RFT53.5152.8581.3760.1561.0066.0053.7751.6186.2887.3485.2373.99
Weighted RFT45.6150.2676.4755.5064.0065.0053.2752.3185.7088.4682.9372.91
Curriculum GRPO45.6152.3387.2559.1761.0070.5059.5054.7886.8389.0784.5978.18
CF-GRPO (ours)48.2559.5982.3562.1064.5073.0058.7556.3589.1792.8085.5378.90

Table R1: Comparison with non-RL and curriculum baselines trained on the same data as CF-GRPO.

W2. The work makes a solid methodological contribution, supported by strong experiments. Could you further clarify its conceptual novelty beyond integrating existing ideas into GRPO?

Thank you for the positive assessment of our methodological contribution and experiments. CF-GRPO’s conceptual contribution is a unified theoretical framework that explains both the failure of outcome-only credit assignment and the complementary roles of our interventions.

  1. A common underlying formulation. While prior approaches address individual aspects of visual grounding through separate interventions, we formalize source-blind credit assignment: under our analysis setting, outcome-reward GRPO reinforces correctness without distinguishing whether it depends on visual evidence.
  2. A principled basis for complementary interventions. Within the same analysis, we show how the three mechanisms control update direction, reinforcement magnitude, and within-group credit allocation, respectively. This establishes why they address distinct aspects of the same problem and how their effects complement one another.

Thank you for your response. Videohallu’s Curriculum RL is a classic and good baseline to compare with . The biggest issue in the paper is too few existing good methods to compare with. Visual hallucinations especially in video is a hard problem. Please add this baseline into your main paper or next version. It will make the paper better. I’ll raise my score, good luck.

Thank you for your positive feedback and constructive suggestions. We agree that comparisons with strong existing methods are important for evaluating progress on video hallucination. In the revised manuscript, we will include the VideoHallu-inspired curriculum GRPO baseline, together with RFT and weighted RFT, in the main experimental section, and clarify its connection to VideoHallu and implementation details. We appreciate your suggestions, which have helped strengthen the paper.

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

This paper proposes CF-GRPO, a reinforcement learning framework designed to reduce video hallucinations by addressing source-blind credit assignment in standard GRPO. The method introduces three complementary mechanisms: counterfactual pair optimization to reduce answer bias, blind-failure advantage shaping to prioritize samples requiring visual information, and an evidence-consistency reward to encourage evidence-aligned reasoning. Experiments on three video hallucination benchmarks demonstrate improvements over standard GRPO while largely preserving general video understanding capabilities.

Critical strengths
  1. Meaningful research motivation. The paper addresses an important limitation of outcome-reward GRPO, highlighting its inability to distinguish visually grounded predictions from correct answers driven by language priors.

  2. Well-structured methodological design. CF-GRPO introduces three complementary mechanisms targeting pair-level updates, sample-level advantage weighting, and within-group reward allocation. The theoretical analysis provides useful motivation for these designs.

Critical weaknesses
  1. Blind-failure advantage shaping does not reliably measure visual grounding. The proposed blind-failure weight assumes that samples answered correctly without visual input should receive less reinforcement. However, the ability to answer correctly without a video does not necessarily imply that the model ignores visual evidence when the video is available. Conversely, failing without visual input does not guarantee that a correct prediction with the video is visually grounded. Therefore, blind-failure probability primarily measures task solvability without visual input rather than the actual reliance on visual evidence. This discrepancy raises concerns about whether the proposed advantage shaping achieves its intended goal.
  2. The evidence-consistency reward may encourage keyword matching rather than faithful visual reasoning. The proposed reward determines evidence consistency by checking whether generated reasoning contains predefined evidence keywords. However, mentioning relevant keywords does not necessarily indicate that the model has correctly perceived or utilized the corresponding visual evidence. In particular, some evidence keywords are derived directly from the queried events and may already appear in the input prompts. Although the authors condition the reward on answer correctness and penalize incorrect answers containing evidence mentions, these mechanisms do not completely eliminate the possibility of rewarding ungrounded reasoning.
  3. Insufficient distinction between counterfactual pair optimization and conventional data augmentation. The proposed counterfactual pair optimization does not introduce an explicit pairwise optimization objective, as each prompt still receives its own GRPO advantage and policy update. The theoretical cancellation of common answer-shift updates also relies on the assumption that paired prompts have similar answer distributions, which is not sufficiently verified empirically. Consequently, the observed improvements may primarily result from balanced positive-negative supervision and the introduction of challenging counterfactual examples rather than the proposed pair-level optimization mechanism. A critical missing comparison is standard GRPO trained on exactly the same counterfactual samples without explicit pairwise grouping.
  4. Insufficient evidence that the performance improvements result from stronger visual grounding. The paper evaluates visual dependence primarily by replacing the video with zero tensors and measuring whether predictions remain unchanged. However, changes in model predictions under blank visual inputs do not necessarily establish that the original predictions relied on correct visual evidence. Such changes may also reflect distribution shifts or altered answer biases caused by removing the video. Although the reported results suggest reduced reliance on language priors, they do not conclusively demonstrate improved visual grounding. Evaluations using controlled visual counterfactuals or independently verified evidence attribution would provide more convincing support for the paper's central claim.
Questions for authors

See weakness

5 comments

We thank the reviewer for the constructive comments and for acknowledging that our work addresses an important limitation of outcome-reward GRPO and that the three mechanisms are complementary and well structured. We respond to each weakness below and include the requested experiments.

W1. Can the blind-failure probability achieve the goal of visually grounded answers?

Yes, the advantage shaping achieves its intended goal.

(1) The blind-failure probability is not designed to measure the reliance on visual evidence. It is used to stop training from reinforcing answers that come from the language prior. For each training question qq, we generate four responses without the video and define g(q)g(q) as the fraction of them that are incorrect. A low g(q)g(q) means that the model can already answer qq correctly without the video, so the outcome reward cannot tell whether a correct answer comes from the video or from the language prior. We therefore multiply the advantages of qq by g(q)g(q), so that such questions contribute little to the update.

(2) To verify this, we train CF-GRPO without the shaping and conduct an evidence attribution evaluation on both models, whose results are reported in Table R1. Each entry is the proportion of questions answered correctly with verified evidence that goes beyond restating the question. The evidence is verified against the video frames by DeepSeek-V4.1-Flash and re-examined by Claude Opus 5.5. With the shaping, CF-GRPO obtains higher scores on most metrics, indicating that the shaping leads to more answers that are grounded in the video.

EventHallusionVideoHallucerVidHalluc
MethodEntireMixMisl.OverallObj.Sem.Ext.OverallAccYesNoACH
CF-GRPO without blind-failure advantage shaping31.2244.5051.5642.5612.5013.2831.8220.8267.3980.0654.7244.30
CF-GRPO (ours)35.0949.4854.9046.8115.0817.0039.7023.2069.6683.9555.3845.17

Table R1: Proportion of questions answered correctly with verified evidence for CF-GRPO with and without blind-failure advantage shaping.

W2. Since some evidence keywords already appear in the input prompts, does the evidence-consistency reward encourage faithful visual reasoning or only keyword matching?

The evidence-consistency reward encourages visual reasoning rather than only keyword matching, which we show in two ways.

(1) The two prompts of a counterfactual pair share the video and the template but differ in the queried event and have opposite answers, so the evidence key in each prompt does not determine the answer. Since the reward credits a correct completion that states the evidence key and lowers the credit of a wrong completion that states it, stating the key is rewarded only together with a correct answer.

(2) We replace the video of every question with a different video from the same benchmark, annotate the new answer with DeepSeek-V4.1-Flash, and re-examine every “Yes” annotation with Claude Opus 5.5. As shown in Tables R2, when a question counts only if it is answered correctly on both videos, CF-GRPO achieves the highest overall score among the trained models on all three benchmarks, indicating that its answers depend on the video rather than on the evidence keys.

EventHallusionVideoHallucerVidHalluc
MethodEntireMixMisl.OverallObj.Sem.Ext.OverallAccYesNoACH
CF-GRPO without evidence-consistency reward41.5546.3277.1553.1657.8264.3145.7345.1882.8789.3876.3569.17
CF-GRPO (ours)46.3649.1280.3956.6661.4667.7146.1947.8686.6390.1983.0874.83

Table R2: Proportion of questions answered correctly both before and after the video replacement for CF-GRPO with and without the evidence-consistency reward.

W3. Does counterfactual pair optimization require a pairwise objective or grouping, how does it compare with standard GRPO on the same samples without grouping, and is the similar-answer-distribution assumption verified?

(1) Counterfactual pair optimization requires neither a pairwise objective nor pairwise grouping. Each prompt is trained as its own group, and the pairing lies in the data. Each counterfactual prompt has a counterpart with the same video and question template that asks about a different plausible event with the opposite answer. A language prior that favors the same answer on both prompts is therefore correct on only one of them, and only the video can make both answers correct. Both prompts of a pair are in the training data and are sampled with equal probability, and we will correct the statement that they are always trained together.

(2) We add the requested comparison in Table R3. Standard GRPO is trained on the same data as CF-GRPO, including the counterfactual pairs, from the same initialization and with the same training budget. CF-GRPO obtains higher Overall scores on EventHallusion and VideoHallucer and higher ACH on VidHalluc.

EventHallusionVideoHallucerVidHalluc
MethodEntireMixMisl.OverallObj.Sem.Ext.OverallAccYesNoACH
GRPO40.3556.4881.3758.1962.0073.5053.5052.4489.0593.9084.2078.78
GRPO + counterfactual data45.6156.9979.4159.4168.5071.0060.0054.6087.6094.4480.7575.95
CF-GRPO (ours)48.2559.5982.3562.1064.5073.0058.7556.3589.1792.8085.5378.90

Table R3: Comparison with GRPO trained on data from the same counterfactual data construction. All rows share the initialization and the training budget.

W4. Can controlled visual counterfactuals or evidence attribution support the paper’s central claim?

Yes, both evidence attribution and controlled visual counterfactuals support our central claim. We thank the reviewer for this suggestion and conduct both evaluations, in which CF-GRPO relies more on the video than the baselines do.

(1) For evidence attribution, DeepSeek-V4-Flash extracts the visual statement on which each answer rests, DeepSeek-V4.1-Flash checks it against the video frames, and Claude Opus 5.5 re-examines every statement judged to contradict the frames. Each entry in Table R4 is the proportion of questions answered correctly with verified evidence that goes beyond restating the question. CF-GRPO obtains the highest Overall score on EventHallusion and VideoHallucer and the highest ACH on VidHalluc.

EventHallusionVideoHallucerVidHalluc
MethodEntireMixMisl.OverallObj.Sem.Ext.OverallAccYesNoACH
SFT34.2144.5653.9244.0111.5619.5028.1420.0567.7979.6855.9044.67
GRPO28.0748.1951.9643.5214.0714.5029.8019.4968.3982.2954.4944.62
OMNEX-VL28.9545.6050.9842.3011.0612.0032.6619.4467.8481.2454.4544.29
CF-GRPO (ours)35.0949.4854.9046.8115.0817.0039.7023.2069.6683.9555.3845.17

Table R4: Proportion of questions answered correctly with verified evidence.

(2) For controlled visual counterfactuals, we replace the video of every question with a different video from the same benchmark. The answer for the new video is annotated by DeepSeek-V4.1-Flash and re-examined by Claude Opus 5.5. Each entry in Table R5 is the proportion of questions answered correctly both before and after the replacement. CF-GRPO again obtains the highest overall scores among the trained models on all three benchmarks.

EventHallusionVideoHallucerVidHalluc
MethodEntireMixMisl.OverallObj.Sem.Ext.OverallAccYesNoACH
SFT45.4533.3372.5547.2658.2558.8547.7243.7582.5185.9479.0967.59
GRPO37.2743.2780.3951.4460.8267.1939.2943.7486.3590.9581.7674.48
OMNEX-VL40.0044.4476.4751.7056.1963.0248.3145.4583.5589.8977.2269.49
CF-GRPO (ours)46.3649.1280.3956.6661.4667.7146.1947.8686.6390.1983.0874.83

Table R5: Proportion of questions answered correctly both before and after the video replacement.

Reviewer_iFQ3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Counterfactual GRPO (CF-GRPO), a method to reduce hallucinations in video understanding in vision language models by modifying the GRPO reward function. The modifications are motivated by the observation that standard GRPO rewards the model for answer correctness but does not account for whether the visual evidence supports the correctness. The authors construct three kinds of counterfactual scenarios - query text pairs having similarity (language prior) but differing responses, video blind samples testing whether the model actually uses evidence from the video, and key evidence samples testing whether the response includes key evidence from the video in the reasoning. GRPO modifications are made ensuring that only the the reward accounts for the counterfactuals without introducing any additional biases or degrading existing performance.

The authors compare CF-GRPO on fine tuning Qwen2.5-VL-7B across video hallucination benchmarks and show improvements over baselines. Additional results show that CF-GRPO maintains performance on general video understanding benchmarks, indicating that performance does not suffer on non-hallucination oriented tasks. Finally, ablations are provided to support the role of each component.

Overall my assessment was that this paper is extremely tedious to read and understand, requiring a significant amount of effort to parse the text. I also had some concerns about the results, described below, that bias my judgement toward rejection.

Edit: Given the clarifications of the authors during the rebuttal, I have raised my score.

Critical strengths
  1. The reward function design modifications are well motivated, with the authors providing both textual and theoretical justification on them.
  2. Despite multiple modifications, there is no performance degradation on non-hallucination tasks.
Critical weaknesses
  1. Concerns about results: Table 1 has the CF-GRPO results bolded indicating it being the best method, but for EventHallusion, the results for the baseline Qwen model seem better across multiple domensions.
  2. Generalization beyond Qwen: All experiments are carried out on Qwen2.5-VL-7B, so it is unclear whether this style of counterfactual GRPO modifications works across models.
Questions for authors
  1. Are the baseline comparisons in Table 1 correct? Which components are actually improved by CF-GRPO and how does that affect the overall score?
  2. Is there any evidence that this method would generalize beyond Qwen2.5-VL-7B?
3 comments

We thank the reviewer for the valuable comments and for acknowledging that the modifications to the reward design are well motivated. We will improve the presentation of the paper in the revision and address each of your concerns point by point below.

W1. Are the baseline comparisons in Table 1 correct? Why does Qwen2.5-VL-7B outperform CF-GRPO on some subsets? How does each component improve the CF-GRPO?

Yes, the comparisons are correct. Table R1 repeats the Qwen2.5-VL-7B and CF-GRPO rows of Table 1. Qwen2.5-VL-7B is biased toward answering “No”. As a result, it outperforms CF-GRPO only on the subsets whose correct answers are all “No” (Entire and Misleading of EventHallusion, and “No” of VidHalluc). CF-GRPO addresses this bias and achieves higher overall scores on all three benchmarks.

As already analyzed in Section 4.2 (lines 370–374), always answering “No” achieves perfect accuracy on these subsets without viewing the video, so higher scores alone do not establish better visual discrimination. Qwen2.5-VL-7B’s bias toward “No” becomes a weakness on EventHallusion’s Mix subset, where 71% of questions are labeled “Yes”: it scores only 37.31, compared with CF-GRPO’s 59.59. CF-GRPO performs well on both answer classes. On VideoHallucer, its positive and negative accuracies are 76.04% and 75.26%, respectively (lines 369–371), supporting that its overall gains reflect improved discrimination rather than a preference for either answer.

Counterfactual pairs make the language prior alone insufficient to answer correctly. Blind-failure advantage shaping stops prompts that the prior already answers from earning reinforcement. The evidence-consistency reward stops rewarding correct answers that the reasoning does not support.

EventHallusionVideoHallucerVidHalluc
MethodEntireMixMisl.OverallObj.Sem.Ext.OverallAccYesNoACH
Qwen2.5-VL-7B64.0437.3195.1059.1743.0053.0043.5046.4882.6873.0492.3367.42
CF-GRPO (ours)48.2559.5982.3562.1064.5073.0058.7556.3589.1792.8085.5378.90

Table R1: Comparison between Qwen2.5-VL-7B and CF-GRPO on the hallucination benchmarks.

W2. Does CF-GRPO generalize to models beyond Qwen2.5-VL-7B?

Yes. To further validate generalization beyond Qwen2.5-VL-7B, we provide additional experiments on Qwen3-VL-8B-Instruct and LLaVA-OneVision-7B, covering both a newer Qwen backbone and a different model family. We use the same training and evaluation settings as in the original experiments. As shown in Table R2, CF-GRPO achieves the best overall score on all three benchmarks with both backbones. We will include these additional results in the revised paper.

EventHallusionVideoHallucerVidHalluc
BackboneMethodEntireMixMisl.OverallObj.Sem.Ext.OverallAccYesNoACH
Qwen3-VL-8BBase65.7933.1693.1457.2163.5068.0051.0052.8384.3073.5895.0369.78
SFT53.5148.7085.2959.1757.5061.5063.0055.4985.9781.8890.0673.36
OMNEX-VL56.1443.0192.1658.9260.0061.0064.5055.8084.7275.7893.6770.67
CF-GRPO (ours)46.4963.2181.3763.0865.0063.5074.5058.6588.0393.5182.5676.75
LLaVA-OneVision-7BBase58.7747.6792.1661.8655.0061.5050.0046.2588.3292.7583.8977.22
SFT54.3975.6591.1873.5954.0052.5056.0047.3083.5884.1183.0577.54
OMNEX-VL53.5164.2588.2467.2453.0061.5056.5047.6188.1691.3384.9878.83
CF-GRPO (ours)60.5374.0993.1475.0657.0061.0059.0048.3788.9189.4388.3979.60

Table R2: Results with other backbones. Bold marks the best result among the trained models for each backbone.

Reviewer_iFQ3OpenReview ↗

Response to Rebuttal

Thank you to the authors for clarifying the Mix vs Misleading data results, and demonstrating the generalization of CF-GRPO. Given these changes, I have raised my score.

Thank you for reading our response and for raising your score. We are glad that our clarification of the Mix and Misleading results and the additional experiments on other backbones have addressed your concerns. We will include the subset-level analysis and the results on Qwen3-VL-8B and LLaVA-OneVision-7B in the revised paper. Thank you again for your time and constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.