VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning

Dear Reviewer Ubhv,

Many thanks for your constructive and insightful feedback! We are glad that you found our VisionThink propose a novel paradigm. We have revised the paper following your suggestions, and will address both your questions and weaknesses below.

---

Questions

[Q1] Discrepancy in Time Acceleration on OCRBench

Thank you for the insightful question. We clarify this from two aspects:

1. On OCRBench, the upper bound for efficiency is Qwen-RL (1/4), which always uses a 1/4-resolution image, resulting in roughly 50% time savings compared to full-resolution input. In contrast, our method uses 1/4 resolution in ~38% of the cases and requests high-resolution input for the remaining ~62%. Achieving a ~20% overall time reduction under this mixed policy is therefore consistent and reasonable.

2. Additionally, we emphasize that our reported time is real-world GPU runtime, not theoretical FLOPs or token counts. As a result, this may introduce variance from multiple system-level factors, but it more accurately reflects practical deployment scenarios.

All evaluation code and measurement procedures will be released to ensure reproducibility. We welcome continued community oversight.

[Q2] Opposite performance between full-image resolution and mixed-image resolution.

Thank you for the insightful question. In a few benchmarks, we observe that the mixed-resolution outperforms the full-resolution setting. We believe this result arises from two key factors:

1. For certain samples requiring resize, the model processes the same image twice at different resolutions, effectively gaining a dual-view perspective. This may serve as a form of implicit data augmentation, yielding marginal performance gains.

2. More importantly, the diverse sampling during RL training enables the model to explore a broader policy space. This exploration may occasionally yield better policies, especially on specific benchmarks.

We will clarify this observation in the final version. Thanks again for your point it.

[Q3.1] Model sensitivity to the penalty settings

Thanks for your professional questions, and we have discussed this problem in the Appendix Sec C.4 Ablation Study on Penalty Control Threshold. The results show that, within an appropriate range, the model’s behavior is not highly sensitive to the exact threshold value.

Thanks again for your point it out, we will consider move this part into the main paper to make it clearly.

[Q3.2] How do different penalties affect the results on various benchmarks?

In Appendix Figure 3, we show how the penalty influences model behavior. Here, we further provide quantitative results across various benchmarks.

As shown in the table below, we report the performance and inference time of models trained with different penalty ratios. Overall, the trends are consistent with Appendix Figure 3: as the penalty ratio increases, the model becomes more inclined to resize the input image, leading to higher inference time.

In general VQA scenarios, model performance does not significantly improve with increased inference time. However, for fine-grained understanding scenarios (e.g., ChartQA, OCRBench), increasing the penalty ratio encourages more image resizing, which results in both longer inference time and improved performance.

Thank you for your constructive question. We will include these results in the paper, as we believe they significantly enhance the completeness and depth of our work.

---

Weaknesses

[W1] Advantage comparison to zoom-in strategy

Thank you for the professional question. We see the advantages of our approach over zoom-in strategies from two key perspectives:

1. Zoom-in methods require strong localization capabilities, which current models like Qwen2.5VL-7B do not fully possess. Inaccurate localization can degrade performance.

2. Zoom-in may require multiple iterative steps due to imprecise region proposals, potentially leading to higher total token consumption than our strategy.

However, we also agree this is a very promising direction and have already noted it as part of our future work (Line 285). We will incorporate this discussion into the camera-ready version to better contextualize the comparison.

[W2] Impact of benchmark characteristics on method behavior

Settings. We selected the OCR-related benchmark ChartQA, along with two benchmarks representing more general VQA scenarios: MME and RealWorldQA. For each, we computed the ratio of samples that triggered image resizing under different input token ranges.

Results. As shown in the table below, there is no clear correlation between image size (token range) and the likelihood of image resizing. Instead, the task type itself has a stronger influence on the model’s behavior. For example, ChartQA consistently exhibits higher image-resize ratios across all ranges, whereas RealWorldQA remains low. This suggests that task characteristics, rather than resolution alone, play a dominant role in guiding VisionThink’s resizing decisions.

Thank you again for your suggestion. We will include this analysis in the camera-ready version of the paper to make our discussion more complete.

[W3] Efficiency on high-resolution images

Thanks for pointing this out. Following your suggestion, we conducted further analysis. As shown in the table below, we focus on DocVQA, the benchmark with the highest image resolution. Our results show that VisionThink achieves significant acceleration while maintaining comparable accuracy.

We believe the reason is that, for high-resolution images, even after resizing (e.g., halving both height and width), the image remains sufficiently clear to solve most questions — even on fine-grained tasks like DocVQA.

We will clarify this point in the camera-ready version. Thank you again for the helpful suggestion.

[W4] Further comparison with additional efficient VLMs

Thank you for the valuable suggestion. In Appendix B.6 (Comparison with Previous Efficient VLMs), we have included additional comparisons with recent state-of-the-art efficient VLMs. We appreciate your reminder and will move this section into the main paper in the camera-ready version to make the comparison more prominent and accessible to readers.

---

Thank you again for helping us improve the paper and hope our response can resolve your concerns! Please let us know if you have any further questions. We will be actively available until the end of rebuttal period. If you feel your concerns are addressed, please consider reevaluating our work. Looking forward to hearing from you :-) !

Dear Reviewer maK1,

Many thanks for your constructive and responsible feedback! We are glad that you found our VisionThink is novel and valuable. We have revised the paper following your suggestions, and will address your questions below first and then address other weaknesses.

---

Questions

[Q1] 1.1 Line87: Overlooked the number of output tokens.

Thank you for the insightful comment. We agree that inference time generally consists of two components: prefilling and decoding, and the number of output tokens impacts the latter. However, we would like to clarify the following:

(1) In contrast to LLMs, VLMs typically generate much shorter textual outputs, often only tens to a hundred tokens. This is significantly smaller by orders of magnitude compared to the number of visual input tokens (which can reach hundreds or thousands). As a result, prefill time dominates total inference time in most VLM scenarios.

(2) Moreover, the inference time reported in our paper is measured in real-world environments, reflecting actual efficiency. As shown in Figure 4, our method achieves speed improvements across multiple benchmarks, highlighting its practical advantage.

(3) We appreciate your insightful suggestion and will incorporate a detailed discussion of this issue in Section 2.2 Computation Complexity to clarify the relative contribution of output token length and better support our claims.

[Q2] Line 46: Explanation for the method name VisionThink

Thank you for the suggestion. Our method is the “think with images” paradigm, where the model first reasons over a coarse visual input and then decides, via reinforcement learning, whether higher-resolution visual information is necessary. Besides, the RL method is also widely used to strengthen the LLM's thinking ability. Therefore, we name our approach VisionThink to emphasize this think with images mechanism.

We appreciate your kind comment and will revise the manuscript to include a clearer explanation, such as:

“We name our approach VisionThink to highlight its think‑then‑look‑clearer paradigm: the policy reasons over a coarse view before deciding whether higher‑resolution evidence is necessary.”

[Q3] 1.2–Line121: Clarification the 130K Dataset.

Thanks for pointing this, and we will detail more in our manuscript. All the datasets are utilized to trained our model via reinforcement learning (Main paper Line122 & Appendix B.5 Line201-205). And we commit to open-source our training and validation datasets, along with the complete codebase, in our GitHub repository.

[Q4.1] Weakness 2.1: Comparison with VisonZip.

Thanks for your interest in VisionZip, you mentioned 99.1% and 98.4% performance retention rates were obtained on LLaVA-1.5. When applying the official github/VisionZip/Qwen_2_5VL procedure to Qwen-2.5-VL, performance drops significantly. As shown in Appendix Table 5, with 50% token retention, VisionZip achieves only 92.0% on ChartQA and 86.5% on OCRBench. In contrast, on these fine-grained benchmarks, our VisionThink maintains 100% and 99.1%, respectively.

Appendix. Table 5: Comparison with Previous Efficient VLM Methods. VisionZip‡ represents the SFT finetuned model.

Method	ChartQA	OCRBench	DocVQA	MME	MMVet	RealWorldQA	POPE	MathVista	MathVerse	Avg.
Vanilla	79.8	81.5	95.1	2316	61.6	68.6	86.7	68.2	46.3	100%
(%)	100%	100%	100%	100%	100%	100%	100%	100%	100%
Down-Sample	62.9	68.8	94.3	2270	54.5	68.8	82.8	62.2	43.1	92.1%
(%)	78.8%	84.4%	99.1%	98.0%	88.5%	100.3%	95.5%	91.2%	93.1%
FastV [7] (ECCV 2024)	72.6	75.8	93.6	2308	52.8	68.8	84.7	63.7	45.0	95.8%
(%)	91.0%	93.0%	98.4%	99.6%	85.7%	100.3%	97.7%	93.4%	97.2%
SparseVLM [72] (ICML 2025)	73.2	75.6	66.8	2282	51.5	68.4	85.5	66.6	45.1	92.2%
(%)	91.7%	92.7%	70.2%	98.5%	83.6%	99.7%	98.6%	97.6%	97.4%
VisionZip [58] (CVPR 2025)	73.4	70.5	93.8	2209	57.0	68.6	86.3	63.7	45.1	95.0%
(%)	92.0%	86.5%	98.6%	95.4%	92.5%	100%	99.5%	93.4%	97.4%
VisionZip‡ [58] (CVPR 2025)	77.3	70.3	93.8	2244	50.1	69.2	91.2	63.1	39.4	95.0%
(%)	96.9%	95.6%	98.7%	96.3%	81.3%	100.9%	107.5%	92.5%	85.1%
VisionThink	79.8	80.8	94.4	2400	68.5	67.1	86.0	67.5	48.0	101.4%
(%)	100.0%	99.1%	99.3%	103.6%	111.2%	97.8%	99.2%	99.0%	103.7%

[Q4.2] Weakness 2.2: Comparison with SFT

Thank you for this professional question. We have also carefully considered this issue and discussed it in Appendix B.6 and Appendix C.

(1) As suggested, we trained a SFT model directly on the RL dataset, incorporating the VisionZip mechanism. As shown in Appendix Table 5 (VisionZip‡), this SFT model does not outperform the training-free version of VisionZip. We attribute this to the limited diversity and coverage of the RL data used for fine-tuning, which is much narrower than the large-scale, high-quality supervised data used by the official Qwen team (Lines 248–253).

(2) In Appendix C.1, we further justify our choice of RL over SFT. Empirically, the RL-trained policy demonstrates better resolution control: it learns when to answer directly and when to request higher-resolution input. In contrast, the SFT model triggers image resizing far more frequently across all benchmarks, indicating weaker policy learning (Lines 267–269).

We will revise the main text to better surface these findings and clarify the rationale for using RL in our method.

[Q4.3] Weakness 2.3: Comparison with keyword-based model / token selection

Thank you for the constructive suggestion. We address this concern from two perspectives:

• In real-world deployment, it is infeasible to accurately classify each input sample based solely on keywords. Whether a high-resolution image is necessary depends not only on the question type but also on the specific visual content. For instance, in the last example of Appendix Section D (Qualitative Results), the question “Are there any fire hydrants here?” contains no OCR- or ChartQA-related keywords, yet resolving it still requires high-resolution evidence.

• In contrast, our VisionThink model is end-to-end and data-driven. It automatically learns a resolution policy conditioned on both the image and the question, allowing it to adaptively reduce image token usage without compromising performance.

We will emphasize this comparison in the revised manuscript to better highlight the advantages of our approach over the keyword-based strategies.

[Q5.1] Weakness 3.1: More QA-pairs examples

Thank you for the suggestion. We have included several qualitative examples in Appendix Section D to illustrate our model’s behavior. We will further enrich this section with additional case studies in the camera-ready version to provide a more comprehensive view of the model’s behaviors.

[Q5.2] Weakness 3.2: Choice of special token for upscaling

Thank you for the insightful question. As discussed in Appendix Section C.3 (Different Prompt Impact), we conducted experiments comparing various prompt designs and special tokens for triggering image upscaling. The results show that following the Qwen official cookbook-style prompt and special token yields the best performance. We will move this analysis into the main paper in the camera-ready version to make the rationale and empirical evidence clearer to readers.

---

Other Weaknesses

[Q6] Writing Suggestions

We sincerely appreciate you for your responsible and detailed review, as well as the constructive writing suggestions. We have carefully revised the paper according to your feedback. Key improvements include:

1. Replaced the term "compress tokens" with "reduction tokens" for better clarity.

2. Refined the motivation section to emphasize the memory and latency constraints caused by long visual token sequences, especially on edge devices.

3. Updated the caption of Figure 3 to:

   Figure 3: Impact of the Penalty Ratio. Applying a penalty to all image-resizing requests or removing the penalty entirely will both lead to model collapse.

4. Moved more details of the Accuracy Reward mechanism to Line 149 for better visibility.

5. Added proper citations for all baselines, e.g.:

   Method	Benchmarks
   SparseVLM [72] (ICML 2025)	—
   FastV [7] (ECCV 2024)	—
   VisionZip [58] (CVPR 2025)	—

6. Clarified throughout the paper that inference time is reported in seconds, not milliseconds or FLOPs.

These revisions will also be reflected in the camera-ready version. Thank you again for your helpful input.

---

Thank you again for helping us improve the paper and hope our response can resolve your concerns! Please let us know if you have any further questions. We will be actively available until the end of rebuttal period. If you feel your concerns are addressed, please consider reevaluating our work. Looking forward to hearing from you :-) !

Reviewer maK1 thanks the authors for their comprehensive and thoughtful rebuttal. I have carefully reviewed all initial reviews and the authors’ responses. The following are additional remarks regarding the rebuttal:

1. The concern raised in W1.1 Line87 is addressed, given that Figure 4 reflects actual efficiency including decoding time, and that the authors will clarify this further in Section 2.2.
2. Appendix Section D provides useful insight into the limitations of rule-based high-resolution selection and clarifies how and when upscaling is triggered by the model. (This reviewer finds this result compelling and strongly recommends including it in the main paper.)
3. It is appreciated that the authors highlighted a direct comparison with VisionZip under the same Qwen2.5-VL baseline.
4. From the perspective of Reviewer maK1, the phrase “think with images (VisionThink)” remains somewhat misaligned with the described behavior “The model first reasons over a coarse visual input and then decides whether higher-resolution visual information is necessary.” Therefore, concern regarding the appropriateness of the name remains. However, this reviewer acknowledges that this is a minor point relative to the main contributions of the work, and is open to further discussion with other reviewers.
5. This reviewer appreciates that the authors listed specific writing revisions. Nonetheless, as mentioned in the initial review, substantial revisions are still expected in the original manuscript.

Based on the authors’ diligent rebuttal, Reviewer maK1 is pleased to raise the score to a 4. Additionally, this reviewer is open to further raising the score through discussion with other reviewers and ACs, particularly regarding points 4 and 5.

Dear Reviewer 2Thu:

Many thanks for your constructive and insightful feedback! We are glad that you found our VisionThink's motivation is clear and reasonable. We have revised the paper following your suggestions, and will address your questions below.

---

[W1 & W2] Table present suggestions

Thank you sincerely for your thoughtful and professional suggestions, we think they significantly improved the clarity of our table presentation.

Following your advice, we revised Table 2 as follows:

• We split the benchmarks into fine-grained and general-scenario categories.
• We removed MMVet and MathVerse as suggested.
• We added mean resolution and token usage per group.
• We included VisionThink‡, a text-only RL baseline trained with full-resolution images using vanilla GRPO and LLM-as-Judge reward.

The revised tables are shown below (Markdown formatting limitations acknowledged):

---

Fine-Grained Benchmarks

---

General-Scenario Benchmarks

---

As shown in the Fine-Grained Benchmarks, VisionThink outperforms prior efficient VLMs by autonomously requesting high-resolution images when necessary. In the General-Scenario Benchmarks, it uses fewer visual tokens while achieving performance comparable to the full-resolution VisionThink‡ baseline. We believe this more clearly demonstrates the effectiveness of our method.

Thank you again for your valuable suggestions. We will further improve the tables in the camera-ready version.

[W3] The Computation of $C_{direct}$ and $C_{high}$ in Eq.5.

The computation of $C_{direct}$ and $C_{high}$ is online. Specifically, $C_{direct}$ and $C_{high}$ represent the number of correct answers from the low- and high-resolution inputs, respectively.

We provide the following pseudo-code for clarity. In actual implementation, the for loop is avoided and replaced with vectorized operations for efficiency.

# Algorithm: Computation of C via VisionThink Inference

# Inputs:
#   inputs             : input data for VisionThink
#   visionthink_model  : model with .inference(inputs) → (outputs, rounds)
#   reward_model       : function outputs → rewards ∈ {0, 1}

# Outputs:
#   C_direct : count of correct answers from low-resolution pass (round == 1)
#   C_high   : count of correct answers from high-resolution pass (round == 2)

# Step 1: Perform model inference
outputs, round_numbers = visionthink_model.inference(inputs)  # rounds ∈ {1, 2}

# Step 2: Compute binary rewards
rewards = reward_model(outputs)  # Shape: (N,), rewards ∈ {0, 1}

# Step 3: Count reward-aligned responses per resolution
C_direct = 0
C_high = 0

for i in range(len(rewards)):
    if rewards[i] == 1 and round_numbers[i] == 1: # correct answers from low‑resolution pass
        C_direct += 1
    elif rewards[i] == 1 and round_numbers[i] == 2: # correct answers from high‑resolution pass 
        C_high += 1

[W4] Data source clarification.

Our datasets are primarily filtered from open-source data, including the LLaVA-OneVision and Cambrian-1 datasets. All data used in our work are publicly available. Furthermore, we commit to releasing our training and validation datasets, along with the complete codebase, in our GitHub repository for full transparency and reproducibility. We welcome continued community oversight.

---

Thank you again for helping us improve the paper and hope our response can resolve your concerns! Please let us know if you have any further questions. We will be actively available until the end of rebuttal period. If you feel your concerns are addressed, please consider reevaluating our work. Looking forward to hearing from you :-) !

Dear Reviewer 5eUF,

Many thanks for your constructive and insightful feedback! We're glad you found the motivation clear, the experiments sufficient, and the paper easy to read. We have revised the paper following your suggestions, and address your questions below.

---

[Q1.1 / W1] More detailed discussion on the resolution issue.

In Figure 1, image resolution is the only controlled variable. Specifically, we resize the image resolution to reduce the number of visual tokens and evaluate the performance of the official Qwen2.5-VL on several benchmarks. The results show that in most general VQA scenarios, even reducing tokens by 75% has minimal impact on the model’s performance. However, in scenarios requiring detailed understanding and OCR-related capabilities, reducing the number of visual tokens leads to a significant drop in performance. This observation suggests that significant token redundancy exists in most scenarios, but a uniform token reduction ratio should not be applied across all tasks.

Thank you for your kind suggestion; we will include the detailed discussion in Paragraph 3, Lines 33–43.

[Q1.2 / W1] Results comparison for same resolution during training and testing.

In Figure 1, which shows the motivation of our paper, the model used is Qwen2.5-VL, which we did not train. For a fair comparison, the results reported in Figure 4, Qwen-RL and Qwen-RL (1/4)—are based on models trained and tested at the same corresponding resolutions. In other words, the training and testing resolutions are consistent.

Thank you for your comments. We will add further explanation in the paper to improve clarity.

[Q2 / W2] VisionThink on stronger visual perception tasks.

Thank you for your insightful suggestion. To evaluate VisionThink on tasks requiring stronger visual perception, we include an additional counting benchmark in our analysis.

Settings. We adopt the widely used CV-Bench, introduced in Cambrian-1 [34], and follow its official setting and prompt for the counting task.

Models. We do not introduce any additional data for task-specific training. All models used are same to those in the main paper. Here, VisionThink‡ denotes the version trained on general VQA tasks with full-resolution inputs using the LLM-as-Judge strategy.

Results. As shown below, both VisionThink and VisionThink‡ outperform the base model (Qwen2.5VL-7B) on the counting benchmark, demonstrating that our approach retains strong performance even on stronger visual perception tasks:

We will include this result in the final version of the paper to better reflect the generalization of our method. Thank you again for the valuable suggestion.

[Q3 / W3] Value of VisionThink compared with choose Qwen-RL or Qwen-RL (1/4) depending on the task type.

Thanks for your constructive question, we want to claim our method value in three aspects:

(1) Balanced Efficiency and Effectiveness

• Qwen-RL can ensure strong performance, but it consumes a significant amount of time even on simple tasks.
• Qwen-RL (1/4) significantly reduces the time cost, but its performance drops sharply on tasks that require detailed understanding, such as OCR-related tasks.
• Our VisionThink achieves performance comparable to Qwen-RL while drastically reducing time consumption—by up to 2×. Compared to Qwen-RL (1/4), VisionThink delivers significantly better performance with similar time efficiency.

(2) Unpredictable Task Types

• In real-world deployment scenarios, we cannot predict in advance what specific task type the user per-sample input belongs to.
• Whether a full-resolution image is needed depends jointly on both the image content and the question type. For example, as shown in Supplementary Section D (Qualitative Results), even for the same image, different questions may require different numbers of image tokens.
• Our VisionThink model is end-to-end and can automatically determine whether more image tokens are needed or if 1/4 of the tokens are sufficient, thereby minimizing image token usage while maintaining performance.

(3) Research Direction

• We encourage the community to move beyond the pursuit of traditional Efficient VLMs and more focus on: Efficient Reasoning VLMs.

To further assess the value of VisionThink, we compare it against a keyword-based resolution selection approach.

Keyword Construction. We use GPT-4o to generate 100 single-word fine-grained/OCR-related keywords (e.g., counting, value, locate) and 100 short phrases (e.g., how many, fine detail).

Setup. The system defaults to Qwen-RL (1/4) for efficient inference. When a keyword is detected in the question, it switches to full-resolution inference via Qwen-RL. This simulates a keyword-triggered token selection policy.

Results. As shown below, due to the diversity in question phrasing, keyword-based strategies generalize poorly and result in suboptimal performance. Moreover, in real-world deployment, VisionThink only requires deploying a single model, while keyword-based approaches require maintaining two separate models, leading to increased resource consumption.

We will emphasize this discussion in the camera-ready version to better illustrate the advantages of VisionThink, which we believe will further clarify the contributions of our paper. Thank you again for your suggestion.

[34] Tong P, Brown E, Wu P, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms[J]. Advances in Neural Information Processing Systems, 2024, 37: 87310-87356.

---

Thank you again for helping us improve the paper and hope our response can resolve your concerns! Please let us know if you have any further questions. We will be actively available until the end of rebuttal period. If you feel your concerns are addressed, please consider reevaluating our work. Looking forward to hearing from you :-) !

Dear Reviewer wk5x:

Many thanks for your constructive and insightful feedback! We are glad that you found this paper's topic is very important and our method has high potential. We have revised the paper following your suggestions, and will address your questions below.

---

[Q1 / W1] Bias Influence of LLM-as-Judge

Your question is very insightful, and we also share concerns about potential biases in the reward model. To mitigate this, we adopt three specific design choices:

(1) Using the LLM-as-Judge not the VLM-as-Judge. Since LLMs generally have stronger capabilities than VLMs, employing an LLM reduces hallucinations and improves reliability.

(2) Filtering the dataset. We filter out subjective open-ended questions that have multiple valid answers, such as image descriptions (Appendix. Line 202–204). The remaining questions have clear ground truth, e.g., Q: <image> Who is the author of this book? A: Dewey Lambdin.

(3) Carefully designing the prompt. Our prompt requires the LLM to return a discrete value: 1 for a correct answer and 0 for an incorrect one, instead of a continuous score. This binary format minimizes ambiguity and reduces the chance of misjudgment (Appendix. Sec. B.1.1).

Furthermore, as shown in the table below, we conduct additional experiments to further investigate this issue. Specifically, we compare the judgments made by GPT-4o and Qwen2.5-72B with those from smaller models such as Qwen2.5-3B and Qwen3-1.7B. While larger models achieve slightly better performance, the smallest model, Qwen3-1.7B, still achieves comparable results under our carefully designed setup. This indicates that LLM model bias has limited influence on our VisionThink.

Thanks again for your professional suggestion, we will include a new subsection titled "The Bias Influence of LLM-as-Judge" in Sec. C (Further Discussions), and provide additional details in Sec. B.1.1 on the prompt design.

[W2] Concern of the dataset source and composition

Our datasets are primarily filtered from open-source data, including the LLaVA-OneVision and Cambrian-1 datasets. All data used in our work are publicly available. Furthermore, we commit to releasing our training and validation datasets, along with the complete codebase, in our GitHub repository for full transparency and reproducibility. We welcome continued community oversight.

[Q2] VisionThink trained with rule-based reward on easily verifiable tasks

Thank you for the valuable suggestion. Following your advice, we trained VisionThink using a rule-based reward signal on tasks where answers can be easily verified.

Datasets. We filtered structured QA samples from our training set where answers can be validated using huggingface/math-verify or string match. These samples are primarily from OCR-related datasets, where answer verification is reliable.

Reward. Instead of relying on the LLM-as-Judge for binary rewards (as in the main paper), we use rule-based verification huggingface/math-verify and string match to provide the reward.

Results. As shown below, the first three models (Qwen-RL, Qwen-RL (1/4), and VisionThink) are same as the main paper that trained with LLM-as-Judge. The final column shows VisionThink (Rule-Based) trained via rule-based reward only. It maintains strong accuracy and efficiency on verifiable tasks like ChartQA and DocVQA.

Discussion. We used the LLM-as-Judge in the main paper to handle general QA scenarios, where reliable rule-based supervision is difficult to define. However, as this experiment shows, for easily verifiable tasks such as OCR-related QA, VisionThink can be effectively trained with rule-based reinforcement learning alone.

Thank you again for the professional suggestion. We will include this experiment in the main paper and provide additional details in Section B.7.

---

Thank you again for helping us improve the paper and hope our response can resolve your concerns! Please let us know if you have any further questions. We will be actively available until the end of rebuttal period. If you feel your concerns are addressed, please consider reevaluating our work. Looking forward to hearing from you :-) !