ViCrit: A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs

We thank Reviewer Nwtd for the encouraging and insightful review. We are pleased that you find the ViCrit task intuitive, the benchmark valuable, and our results strong. Below we address your key concerns and suggestions in detail:

W1: Generalization beyond Qwen2.5-VL

• We confirm that ViCrit generalizes to other model architectures and report results on InternVL-2.5-8B[1]. To assess generalizability, we applied ViCrit training to InternVL-2.5-8B without any architecture-specific tuning. The resulting model, ViCrit-RL-InternVL-2.5-8B, shows consistent improvements across all evaluated benchmarks:

• These results support that ViCrit can transfer across model architectures and training pipelines. We agree that further applications (e.g., to LLaVA-1.5) are valuable, and we plan to explore this in future releases.

W2: Drop in “Text” hallucination performance

• The observed drop in “Text” hallucination accuracy stems from training data imbalance rather than a limitation of the ViCrit objective. As detailed in Section 4.1, our PixMo-Cap-derived training set is dominated by natural images (~59%), with scene-text-heavy and document images comprising only ~24% and ~10% respectively. Since “Text” hallucinations are disproportionately concentrated in these underrepresented categories, models have limited exposure to them during RL fine-tuning.
• Importantly, ViCrit-RL still outperforms the base model across 10 of 11 benchmarks, including strong gains on abstract and document-heavy tasks (e.g., +3.3 on VLMsAreBlind, +6.4 on CharXiv). This suggests that while one hallucination type underperforms, the overall visual perception is significantly improved.
• We have begun exploring class- and domain-aware data resampling and weighted reward shaping to mitigate this issue in ongoing experiments. Additionally, we are expanding the ViCrit training set with more scene-text and document-rich samples (e.g., charts, scanned documents, GUI mockups) to support stronger generalization across all hallucination types.
• We view this as an opportunity to develop domain-adaptive RL pipelines for VLMs, where model sensitivity to hallucination types can be tuned based on application needs (e.g., OCR-critical tasks vs. object-centric ones).

W3: Further SFT baseline for isolating the role of GRPO

• We followed your suggestion and added a new SFT baseline that generates the non-hallucinated caption given the hallucinated one. We refer to this model as Qwen2.5-VL-7B-CapSFT-Generation, and report its performance below:

• This baseline confirms that ViCrit-RL outperforms SFT-based alternatives. These results isolate the value of reinforcement learning and verifiable perception-based rewards as central to ViCrit’s generalization capabilities.

Q1: Impact of caption length

• We appreciate the reviewer’s insightful question—caption length plays a crucial role in ViCrit’s design and difficulty. ViCrit is specifically constructed to promote fine-grained, holistic visual perception by embedding a single subtle hallucination into long, paragraph-length captions (~200+ words). These long contexts force the model to verify a wide range of objects, relations, and attributes before locating the anomaly—closely mimicking real-world multimodal grounding tasks.
• Shorter captions may reduce the perceptual search space and weaken the RL signal, as hallucinations become more easily detectable and less dependent on visual grounding. This risks rewarding shallow pattern-matching strategies, rather than encouraging true image-text alignment.
• Although our current dataset (PixMo-Cap) does not include a sufficient volume of shorter captions for a clean ablation, we plan to conduct a systematic study by synthetically truncating captions or controlling their verbosity. Such an analysis would quantify the tradeoff between perceptual load and learning benefit, and help determine whether longer contexts better support chain-of-thought-style visual reasoning.
• We believe caption length could also be a lever for curriculum-based RL, where models start with shorter, easier examples and progress to dense, ambiguous scenes—an exciting future direction we plan to explore.

We appreciate your suggestions, especially regarding baselines and performance analysis. We will incorporate these findings and discussions in the revised version to further clarify the strengths and limitations of ViCrit.

Reference:

[1] Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling. Chen et. al. arXiv:2412.05271

We thank Reviewer DvR5 for the detailed and thoughtful review. We are encouraged that you find our ViCrit task interesting, our dataset construction reasonable, and our paper clear and well-written. Below we address your questions and concerns:

W1&Q1: What is the advantage of ViCrit over standard VQA datasets?

• ViCrit is designed specifically for reinforcement learning with verifiable perception tasks—something standard VQA datasets cannot support. While VQA datasets provide short question-answer pairs, they rarely require exhaustive scene grounding. In contrast, ViCrit uses long-form captions with one injected hallucination, requiring the model to check all visual details and identify the corrupted span. This makes the task more challenging and better aligned with scalable RL training, where exact match rewards are necessary.
• Standard VQA datasets often admit multiple plausible answers or rely on language priors. For example, "What is the man doing?" may be answered "playing tennis" or "hitting a ball"—both acceptable, but hard to score in RL. ViCrit avoids this ambiguity by offering binary string-matching rewards at the span level, enabling efficient and noise-free optimization.
• We will include a comparison of task properties (e.g., ambiguity, span length, reward verifiability, image domain diversity) with standard VQA datasets in the final version.

W2&Q2: Why are there no comparisons with reasoning models like Vision-R1 or OneVision?

• We now include experimental comparisons against Vision-R1-7B[1], R1-OneVision-7B[2], and ThinkLite-VL-7B[3]. All models use Qwen2.5-VL-7B as the base. The table below compares ViCrit-RL-7B with these strong reasoning-focused models:

• ViCrit-RL-7B achieves the best hallucination mitigation (CHAIRs/i, MMHal), and competitive results on general reasoning benchmarks. It outperforms Vision-R1 and R1-OneVision on all fIVE visual-centric reasoning benchmarks (MMMU, MMStar, MM-Vet, Blind, CharXiv), and matches or exceeds them on MathVista, MathVision, and MathVerse.
• Furthermore, we apply ViCrit-RL on top of ThinkLite-VL-7B, yielding improved performance across the board. This joint model achieves a new SOTA of 75.5 on MathVista among 7B-level VLMs, demonstrating ViCrit's effectiveness and compatibility with existing reasoning-focused methods.

W3&Q3: Why does training on a perception task improve reasoning performance?

• ViCrit improves reasoning by improving visual grounding—especially for math and abstract image tasks. Many reasoning benchmarks (e.g., MathVista, CharXiv, VLMsAreBlind) require the model to perceive layout, count objects, read text, and track spatial relations before reasoning. ViCrit forces models to internalize how to perceive accurately, which in turn supports better reasoning outcomes.

• We illustrate this in Figure 5 of the paper, where ViCrit-RL-72B generates more accurate and complete intermediate thoughts than the base model. These qualitative results show ViCrit-trained models pay more attention to visual details (e.g., color, size, shape, text), which enables more robust chain-of-thought reasoning.

• We will include additional case studies in the appendix to further show how perception improvements propagate to reasoning quality.

We thank the reviewer again for raising these important points. We believe ViCrit fills a critical gap by offering the first vision-centric, verifiable RL task for VLMs, and we are excited to open-source the code and benchmark to support broader adoption and future research.

Reference:

[1] Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models. Huang et. al. arXiv:2503.06749

[2] R1-Onevision: Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization. Yang et. al. arXiv:2503.10615

[3] SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement. Wang et. al. arXiv:2504.07934

We thank Reviewer Fe2P for the thoughtful and constructive review. We are glad that you find our ViCrit task to be novel, practically valuable, and efficient to evaluate, and that you appreciate the diagnostic contribution of ViCrit-Bench across four visual domains. We respond below to your concerns regarding reproducibility, comparisons, and generalizability:

W1: GRPO training details and visualization metrics

• All GRPO hyperparameters are now disclosed, and training reward values are reported to illustrate learning dynamics. Please see the table of GRPO hyperparameters below, along with stepwise reward progression for both 7B and 72B models. Due to NeurIPS rebuttal policy, we cannot include figures, but we will include full training curves in the final version.

Here are all the training hyperparameters used for GRPO training:

ViCrit-RL-7B training reward curve

ViCrit-RL-72B training reward curve

• We will open-source our training code and logs for full reproducibility. These will be made available with our camera-ready version, including configuration files and scripts for ViCrit task construction and GRPO optimization.

W2: Lack of comparisons to other tasks or datasets

• ViCrit is not designed to replace existing tasks but rather complements them with a perception-verifiable RL objective. Our goal is to show that ViCrit enhances fine-grained visual perception via reinforcement learning with a verifiable reward—an orthogonal axis to many existing multimodal tasks that optimize other cognitive capabilities.
• We include new comparisons against SOTA models trained on other tasks (Vision-R1[1], R1-OneVision[2], ThinkLite-VL[3]), showing ViCrit’s strong hallucination mitigation and comparable general reasoning. The results are provided in the following table:ViCrit-RL-7B reduces hallucinations (CHAIRs/i) while retaining competitive accuracy on general tasks like MMVet, Blind, and Charxiv.

• To evaluate compositional benefits, we applied ViCrit-RL on top of ThinkLite-VL-7B and observed further improvements on both hallucination and reasoning benchmarks. Notably, ThinkLite-VL-7B+ViCrit-RL sets a new SOTA on MathVista (75.5) among 7B-level VLMs. This suggests ViCrit training encourages better visual verification and reduces hallucination even when stacked atop other reasoning-optimized models.

W3: Experiments on other base models

• We validated ViCrit on InternVL-2.5-8B[4] to test generality, and observed consistent gains across all evaluated benchmarks. As shown in our results, ViCrit-RL improved both hallucination mitigation and downstream reasoning performance compared to the base InternVL-2.5-8B. This confirms that ViCrit is architecture-agnostic and generalizes well beyond our primary Qwen-2.5-VL base.

Thank you again for your thoughtful review. We are excited to open-source ViCrit and ViCrit-Bench to the community and believe they can serve as foundational tools for future multimodal RL research.

Reference:

[1] Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models. Huang et. al. arXiv:2503.06749

[2] R1-Onevision: Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization. Yang et. al. arXiv:2503.10615

[3] SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement. Wang et. al. arXiv:2504.07934

[4] Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling. Chen et. al. arXiv:2412.05271

Dear Reviewer Fe2P

We sincerely thank you again for your valuable feedback, which has greatly helped us improve the quality of our work. As the discussion phase is now open, we kindly remind you to review our rebuttal. We are happy to discuss more if you have any further concerns. If our responses adequately address your concerns, we kindly ask you to consider adjusting the corresponding scores. Thank you again for your time and effort!

Dear Reviewer Fe2P,

We sincerely thank you again for your valuable feedback, which has greatly helped us improve the quality of our work. As the discussion phase is nearing its end—with only two days remaining—we kindly remind you to take a moment to review our rebuttal.

Two other reviewers have already responded, expressed satisfaction with our clarifications, and one of them accordingly updated the score. If you have any remaining concerns, we would be very happy to engage in further discussion. Otherwise, if our responses have adequately addressed your comments, we kindly ask you to consider adjusting your score as well.

Thank you once again for your time and thoughtful review!

We thank Reviewer qk6D for the thoughtful and constructive feedback. We appreciate that you find our proposed ViCrit task to be novel and useful, and that you recognize the value of our extensive evaluations and the introduction of ViCrit-Bench. Below, we address all concerns and questions raised:

W1: Limited RL training details

• We now provide full GRPO hyperparameters and reward values across training steps to support reproducibility and transparency. The table below lists all GRPO hyperparameters used in our ViCrit-RL training:

• We report reward progression during training as a substitute for full reward curves. Due to rebuttal policy, we cannot include figures or links. Below are reward values at various training steps, we will include the complete training reward curve in the final version of the paper to better illustrate the training dynamics:

ViCrit-RL-7B

ViCrit-RL-72B

W2: Missing training recipe

• We used EasyR1 as our training codebase and will release recipes for reproducibility. We use Qwen-2.5-VL-7B/72B as base models and train on ViCrit directly via GRPO, with no supervised fine-tuning or knowledge distillation. Reward formulation is detailed in Section 3.2, lines 159–165, and we will include the complete training recipe in the final version and code release.

W3: Missing human performance on ViCrit-Bench

• We agree that human accuracy is a crucial baseline, and we have now begun collecting expert annotations to report this in the camera-ready version. Establishing a human upper bound will help contextualize model performance, especially given the fine-grained nature of ViCrit tasks (e.g., subtle hallucinations in shape, count, or text). We are employing annotators familiar with vision-language tasks and providing the full paragraph-length captions and images as shown in ViCrit-Bench, under the same evaluation setting used for models.
• We thank the reviewer for highlighting this point—we believe it will strengthen the benchmark’s utility and interpretability for future work.

W4: Lack of benchmark examples or case studies

• Figure 3 provides examples from each hallucination category; we will expand with full case studies in the appendix. Due to page limits, we only included hallucination spans, not full captions. The revised version will add full examples for each hallucination type to better convey task difficulty and practical use.

Q1: GUI images not considered in domain categorization

• GUI images are included and categorized under the “document” domain. As shown in the second example of Figure 3, our benchmark includes GUI-like screenshots (e.g., software interfaces, forms). We clarify this classification in Section 4.1, lines 175–181, and will make this more explicit in the revised text.

Thank you again for your helpful suggestions. We are committed to improving reproducibility and clarity in the final version and believe ViCrit and ViCrit-Bench will provide valuable tools for advancing VLM research.

Dear Reviewer qk6D

We sincerely thank you again for your valuable feedback, which has greatly helped us improve the quality of our work. As the discussion phase is now open, we kindly remind you to review our rebuttal. We are happy to discuss more if you have any further concerns. If our responses adequately address your concerns, we kindly ask you to consider adjusting the corresponding scores. Thank you again for your time and effort!

Dear Reviewer qk6D,

We sincerely thank you again for your valuable feedback, which has greatly helped us improve the quality of our work. As the discussion phase is nearing its end—with only two days remaining—we kindly remind you to take a moment to review our rebuttal.

Two other reviewers have already responded, expressed satisfaction with our clarifications, and one of them accordingly updated the score. If you have any remaining concerns, we would be very happy to engage in further discussion. Otherwise, if our responses have adequately addressed your comments, we kindly ask you to consider adjusting your score as well.

Thank you once again for your time and thoughtful review!