Thanks for your acknowledgment and interest in our work! We sincerely appreciate your thorough and insightful comments on our work, and we will address each of your main concerns below:

Q1: The experiments primarily focus on a small set of LVLMs.

R1: Thanks for your insightful comment. In our paper, we select a range of LVLMs with diverse architectures and parameter scales, spanning from proprietary models (GPT-4o, GPT-4o-mini) to open-source models (Qwen2-VL, LLaVA-1.5), in order to provide the broadest possible interpretation of the role of visual thoughts.

Q2: The correlation analysis of clarity and conciseness (Figure 6) relies on manual scoring.

R2: Thanks for your constructive comment. We also employ the automatic evaluation method CLIPScore to assess all the visual thoughts presented in Table 1. Specifically, we compute the classification probability between two categories: “This is a clear (annotated logical) image” and “This is not a clear (annotated logical) image” (for G-IMG, content in parentheses is removed). These probabilities serve as quantitative indicators for clarity score of visual thought.

Table 1: Correlation analysis between clip-based clarity scores and model performance on GPT-4o-mini. || Pearson Correlation| P | |-|:-:|:-| | S-LANG | 0.692| 0.06 | | N-LANG | 0.693| 0.04 | | G-IMG | 0.731 | 0.04| | E-IMG | 0.646| 0.06|

From the results, we can observe that the trends from human evaluation and automatic evaluation are consistent.

We will add more discussion in the next version.

Q3: Certain sections lack a detailed discussion of why certain visual thought expressions perform better in specific tasks.

R3: Thanks for your kind feedback. To investigate the performance differences of visual thoughts across various tasks, we analyze them in conjunction with the characteristics of specific task sets. For instance, CoMT—which places greater emphasis on visual information interaction—tends to benefit more from image-form visual thoughts that effectively convey information. In Q6, we further examine the impact of noise within visual thoughts on task-specific performance, thereby further revealing the factors that contribute to the performance gains brought by visual thoughts. We will add more discussion in the next version.

Q4: The experiments for attention and information flow analyses is unclear how generalizable the findings are across different LVLM architectures or scales.

R4: Thanks for your insightful feedback. We follow your suggestion to add more models for evaluation.

1. For attention analysis:

We conduct attention analysis across four models (LLaVA-7B/13B and Qwen2.5-VL-3B/7B) by comparing attention to the input image (w/o visual thoughts) and to visual thoughts (with VT). According to the result, we observe that across most layers—and especially in the latter half—attention to visual thoughts consistently exceeds that to the original image, indicating that visual thoughts can more effectively convey visual information during reasoning, consistent with findings in the paper.

Moreover, we find that models with more layers exhibit stronger attention to visual thoughts, suggesting that architecture depth is a key factor in leveraging visual thoughts.

For methodological details and specific experimental results, please refer to our detailed response to Reviewer #NrVc (R1).

2. For information flow analysis:

We incorporate Qwen2-VL-2B to examine information flow under different types of visual thoughts. According to the experimental results, all variants of visual thoughts yield stronger information flow to reasoning stages than the original image, consistent with our findings in LLaVA models.

Notably, unlike LLaVA, Qwen2-VL exhibits the highest information flow in the n-lang setting, likely due to its stronger language generation capabilities.

For a detailed explanation and discussion, please refer to our response to Reviewer #NrVc (R2).

Due to space constraints, we summarize only the key insights here. We sincerely appreciate your understanding!

We will add more discussion in the next version.

Q5: The finding of 'forcing reasoning solely from the original image can impair performance more than reasoning directly from the query' appears counterintuitive and lacks adequate discussion.

R5: Thanks for your valuable comment. We sincerely believe there may have been a misunderstanding regarding our explanation. Actually, what we intended to convey is that in Figure 4(b), removing the visual thought module can, in some cases, lead to performance even lower than the Query-Only baseline. However, this does not imply that all instances of the “w/o Visual Thought” condition perform worse than the Query-Only baseline (as clarified in the first paragraph of Section 3).

Moreover, this phenomenon is observed only in the Math subset, primarily because many Math questions contain sufficiently detailed textual descriptions that allow the model to perform reasoning without relying heavily on visual features. Removing visual thoughts in such cases disrupts the effective transmission of visual cues, which unexpectedly degrades performance. In contrast, for other datasets, the queries themselves carry minimal information (e.g., questions such as “How many people are in the image?”), and thus the Query-Only condition naturally performs worse than “w/o Visual Thought”.

To further address your concerns, we have included additional experiments, as illustrated in Table 2. Similar trends are observed not only in GPT-4o but also in other models such as GPT-4o-mini and Gemini.

Table 2: Results on w/o visual thought v.s. Query-Only. ||chess|graph|math| |-|:-:|:-:|:-:| |GPT-4o-mini w/o visual thought|6.00%|38.00% |9.00%| |GPT-4o-mini Query-Only|0.00%|3.00%|26.00%| |Gemini-2.0-flash w/o visual thought |4.00%| 24.00%|17.00%| |Gemini-2.0-flash Query-Only|1.00%|0.00%|29.00%|

We will incorporate more detailed discussion on this issue in future versions of the paper to ensure clarity and alleviate any potential confusion.

Q6: How do errors injected from inaccurate segmentations or irrelevant generated images impact reasoning performance, and how are they mitigated?

R6: Thanks for your constructive feedback. To evaluate the potential performance variations in MCoT caused by errors introduced through image-form visual thoughts, we conduct the following experiment:

For E-Img: To evaluate the potential errors introduced by E-Img, we controll the editing effects by adjusting the parameters of the editing tool and observe corresponding changes in MCoT performance. Specifically,

1. We select GroundingDino as the editing tool and manipulate the box threshold and text threshold parameters to vary the accuracy of grounding. Images with poor grounding mistakenly annotate irrelevant objects as grounded, resulting in inaccurate edited images.
2. These edited images are then ranked by GPT-4o based on grounding quality, and the resulting ranking is used to quantify the introduced errors by assigning specific scores to specific ranks.
3. We sample 100 instances from V*Bench-attributes which are sensitive to E-Img and remove the instance failing to perform grounding, using GPT-4o-mini as the model. The results are shown in Table 3.
4. We include the method without incorporating visual thoughts (no-vis) as the baseline.

For G-Img: To facilitate the control of the quality of the G-IMG, we controll the generation effects by adjusting the parameters of the generation models and observe corresponding changes in MCoT performance. Specifically,

1. We adopt the open-sourced Stable Diffusion v1.5. We manipulate the model’s guidance scale parameter to control the degree to which the generated image adheres to the input prompt. Images generated with a low guidance scale often fail to reflect visual information relevant to the original query, thereby resulting in irrelevant generated images.
2. We then used CLIPScore to evaluate the similarity between the generated images and the input prompts, serving as a metric to quantify the introduced error.
3. We sample 100 instances from M3CoT which is sensitive to G-Img as the experiment data, using GPT-4o-mini as the model. The results are presented in Table 4.
4. We include the method without incorporating visual thoughts (no-vis) as the baseline.

Table 3: Evaluation of E-Img error sensitivity. |box threshold & text threshold(↑)|noise score(↓)|acc(↑)| |:-:|:-:|:-:| |0.07 & 0.05| 0.9278|29.17| |0.14 & 0.10| 0.7806|33.33| |0.21 & 0.15| 0.6333| 43.06| |0.28 & 0.20|0.4167|54.17| |0.35 & 0.25|0.2417| 55.56 | |no-vis|-|34.72| Table 4: Evaluation of G-Img error sensitivity | guidance scale(↑) | clipscore(↑) | noise score(↓) | acc(↑) | |:-:|:-:|:-:|:-:| |0|0.6720|0.3280|68.57| |2.5|0.7815|0.2185|73.33 | |5|0.8168|0.1832|74.29| |10|0.8401|0.1599|77.14| |15|0.8519|0.1481|77.14| |no-vis|-|-|72.38|

Based on the results in Tables 3 and 4, we observe a positive correlation between the quality of edited and generative image and the MCoT performance.

We further observe that poor-quality edited images (box threshold & text threshold=0.07 & 0.05) or generated images (guidance scale=0) can even impair the reasoning capabilities of MCoT (lower than no-vis). Through manual inspection, we find that those images contain substantial misleading information (e.g., grounding numerous objects that do not meet the specified conditions or generating images that convey incorrect visual content). This indicates that inaccurate visual thoughts may introduce noise and, in turn, lead to degraded performance.

To mitigate the introduced errors, we can adopt more powerful editing and generative models, and leverage general-purpose LVLMs to evaluate the resulting images for timely correction.

We will DEFINITELY add more the discussion for mitigate your concerns in the next version.

Most of the questions have been addressed, and I am considering updating my final rating. I would like to see these discussions included in the final version.

I have no further questions. While I find the results valuable, the insights provided are somewhat limited. Therefore, I will maintain my current rating.

Thanks for your acknowledgment and interest in our work! We sincerely appreciate your thorough and insightful comments on our work, and we will address each of your main concerns below:

Q1: Conclusions in effectiveness verification of visual thought can be incoherence of the resulting context.

R1: Thank you for your constructive feedback. To address your concern, we compare direct prompting with GPT-4o to generate results (Direct) against explicitly instructing it to avoid generating any visual content (w/o VT prompting). As shown in Table 1, the results obtained by w/o VT prompting indicate a performance drop when visual information is will never be mentioned by model, even when the context is complete.

It also aligns our major conclusion: Visual thoughts are the main visual information carrier.

We will include more detailed discussions in the next version.

Table 1: Results on Direct prompting vs. w/o VT prompting. || CoMT-Selection | CoMT-Update | Math-Breakpoint | |:-:|:-:|:-:|:-:| | Direct | 13.00| 42.00| 84.00| | w/o VT prompting | 12.00| 32.00| 72.00 |

Q2: The authors state that "removing visual thoughts leads to performance worse than using the query alone" maybe not correct.

R2: Thank you for your meticulous and insightful comments. We sincerely believe there may have been a misunderstanding regarding our explanation. Actually, what we intended to convey is that in Figure 4(b), removing the visual thought module can, in some cases, lead to performance even lower than the Query-Only baseline. However, this does not imply that all instances of the “w/o Visual Thought” condition perform worse than the Query-Only baseline (as clarified in the first paragraph of Section 3).

Moreover, this phenomenon is observed only in the Math subset, primarily because many Math questions contain sufficiently detailed textual descriptions that allow the model to perform reasoning without relying heavily on visual features. Removing visual thoughts in such cases disrupts the effective transmission of visual cues, which unexpectedly degrades performance. In contrast, for other datasets, the queries themselves carry minimal information (e.g., questions such as “How many people are in the image?”), and thus the Query-Only condition naturally performs worse than “w/o Visual Thought”.

To further address your concerns, we have included additional experiments, as illustrated in Table 2. Similar trends are observed not only in GPT-4o but also in other models such as GPT-4o-mini and Gemini.

Table 2: Results on w/o visual thought v.s. Query-Only. || chess | graph |math| |-|:-:|:-:|:-:| |GPT-4o-mini w/o visual thought|6.00%|38.00% |9.00%| |GPT-4o-mini Query-Only|0.00%|3.00%|26.00%| |Gemini-2.0-flash w/o visual thought |4.00%| 24.00%|17.00%| |Gemini-2.0-flash Query-Only|1.00%|0.00%|29.00%|

We will incorporate more detailed discussion on this issue in future versions of the paper to ensure clarity and alleviate any potential confusion.

Q3: The problem with the baseline and suggestions of including the vanilla baseline without any CoT reasoning at all.

R3: Thanks for your insightful feedback. The w/o VT of Section 4 and 'w/o visual thought' from Figure 4 represent the same meaning but not the same implementation as not including the visual thoughts in the rationale. Specifically, we implement methods in Section 4 by explicitly prompting the model and we implement methods in Figure 4 by removing the VT manually.

In addition, we follow your suggestion to add the vanilla baseline without any CoT reasoning and the results are shown as:

Table 3: Results on w/o CoT. || MMVP | V*Bench-Position | V*Bench-attributes | M3CoT-physical | M3CoT-social | M3CoT-temporal | CoMT-deletion | CoMT-selection |CoMT-update|AVG| |-|-|-|-|-|-|-|-|-|-|-| | w/o VT | 74.33 | 53.95| 54.78| 88.89 | 76.86 | 79.67 | 26.50|19.50|37.00 |56.83 | | N-LANG |85.33| 57.89 | 63.48 | 88.89 | 78.93 | 83.74| 33.50| 25.50 |37.50|61.64| | S-LANG|84.33|63.16| 64.35| 90.00 | 78.51 | 82.93 | 29.50 | 18.00 |42.00|61.42| | E-IMG | 83.00 | 59.21 | 65.22 | 90.00 | 78.10 | 86.18 | 34.00| 28.50 |50.00 |63.80| | G-IMG| 78.00 | 59.21 | 59.13 | 92.22 | 78.93| 86.18 | 33.50 | 28.50 |46.50 |62.46| | w/o CoT | 77.67 | 57.89| 63.48| 87.78| 76.03 | 58.54 | 32.50 | 21.50 |41.50|57.43|

We observe that the trends of removing Visual Thoughts (VT) and CoT are not consistent. As shown in Table 3, for benchmarks that require more reasoning steps, such as M3CoT, which involves extensive multi-step reasoning, the performance of the w/o CoT variant significantly degrades.

Moreover, in most vision-reasoning scenarios, removing VT leads to even worse performance compared to removing CoT. This further substantiates our claim that visual thoughts are the core of MCoT. Rather than merely providing an additional reasoning process, VT plays a crucial role in conveying visual information throughout the reasoning pipeline.

We will add more discussion in the next version.

Q4: The conclusions drawn from the attention plots are not entirely convincing.

R4: Thank you for your constructive feedback.

Attention Distribution Analysis: To further support the conclusions drawn in Figure 7, we perform additional quantitative analyses to four models with varying layer counts: LLaVA-1.5-7B (32 layers), LLaVA-1.5-13B (40 layers), Qwen2.5-VL-3B (36 layers), and Qwen2.5-VL-7B (28 layers).

1. For each model, we extract internal attention to the original image using the NO-VIS method and compare it to attention directed at visual thoughts using the N-LANG, S-LANG, E-IMG, and G-IMG methods, as shown in Figure 7(b).
2. We then compute the percentage of layers where attention to visual thoughts exceed that to the input image.
3. This percentage is also calculated specifically for the latter 50% of layers (values shown in "()").

Table 4: The proportion of model layers in which the attention associated with visual thoughts surpasses that of the original input image. Values in "()" correspond to results calculated from the latter 50% of the model’s layers. ||Layer |N-LANG(%)|S-LANG(%) |E-IMG(%)|G-IMG(%)| |:-:|:-:|:-:|:-:|:-:|:-:| |LLaVA-1.5-7B|32|43.75 (12.50)|40.63 (6.25)| 96.88 (100.00) | 93.75 (100.00) | |LLaVA-1.5-13B|40|20.00 (0.00)|35.00 (5.00) | 95.00 (100.00) | 92.50 (100.00) | |Qwen2.5-VL-3B| 36 |97.22 (100.00) | 86.11 (100.00) | 72.22 (88.89) | 58.33 (100.00) | |Qwen2.5-VL-7B| 28|100.00 (100.00)| 92.86 (100.00) | 60.71 (100.00) | 60.71 (100.00) |

Results:

• Table 4 shows that in most layers across both LLaVA and Qwen models, attention favors visual thoughts over the original image, suggesting that reasoning relies more on visual thoughts.
• Notably, in the latter half of the layers, attention to visual thoughts often dominates entirely, reaching 100% in some cases, indicating their superior capacity to convey visual information. These findings align with our overall conclusions.

* Under the N-LANG and S-LANG settings for LLaVA, attention to visual thoughts is relatively low, likely due to the model’s limited self-generated language capacity. The resulting captions and scene graphs are overly simplistic, failing to capture visual content adequately.

Table 5: The difference in attention score to visual thoughts relative to the original image across all layers. || Layer | N-LANG | S-LANG| E-IMG | G-IMG| |:-:|:-:|:-:|:-:|:-:|:-:| |LLaVA-1.5-7B| 32 | -0.0529 |-0.0472| 0.0156 | 0.0256| |LLaVA-1.5-13B |40| -0.0481| -0.0275| 0.0552 | 0.0637 | |Qwen2.5-VL-3B|36|0.1280| 0.0438| 0.0229 | 0.0388 | |Qwen2.5-VL-7B|28|0.1069| 0.0553 | 0.0189 | 0.0152|

Average Attention Score Anlaysis: We compute the attention score differences induced by visual thoughts relative to the original image, averaged across all layers. The findings are:

• As shown in Table 5, both LLaVA (7B->13B) and Qwen(7B->3B) exhibit increased attention gains with deeper model layers.
• For Qwen2.5-VL, the 3B variant, despite its smaller parameter size, exhibits a greater attention increase than the 7B variant, likely due to its deeper architecture. This indicates that the influence of visual thoughts is more dependent on model design than parameter count.

We will emphasize this section further and expand the discussion in the next version.

Q5: What metric is plotted on the y-axis of Figure 5? Why do all columns add up to 1? Please clarify in the caption or main text.

R5: Thank you for your kind reminder. The y-axis in Figure 5 represents the normalized relative accuracy improvement of each dataset incorporating visual thoughts. The y-axis metric of $i$-th dataset can be mathematically represented as:

$Metric_i^y = \frac{\delta acc_i}{\sum_{j=1}^{n} \delta acc_j},$

where $\delta acc$ represents the relative accuracy improvement, $n$ represents the total number of datasets. We then calculate this metric for each type of visual thought, which produces the results in Figure 5 in our paper.

We will add the details in the next version.

Q6: Questions about the experiments on disruption of information flow.

R6: Thank you for your valuable suggestion. In the information flow disruption experiment presented in Section 5.2, we adopt the methodology outlined in [1,2], in which the information exchange between designated token pairs is obstructed by assigning a value of $-\infty$ in selected transformer layers.

We will provide more explanation and clarification in the next version.

[1] Label words are anchors: An information flow perspective for understanding in-context learning.

[2] Lifting the Veil on Visual Information Flow in MLLMs: Unlocking Pathways to Faster Inference.

Q7: Typo Problems

R7: Thanks for your valuable feedback. We will follow your suggestions to fix all typo problems in the next version.

Hope our clarification can address your concerns, and we sincerely hope that you can reconsider our work in light of these clarification.

Thanks for your acknowledgment and interest in our work! We sincerely appreciate your thorough and insightful comments on our work, and we will address each of your main concerns below:

Q1: Is the finding in Section 5.1 only appearing in deep models? Is visual thought more beneficial for deeper models?

R1: Thank you for your constructive feedback. To investigate how visual thoughts influence internal attention across models of different depths. We select four models with varying numbers of layers: LLaVA-1.5-7B (32 layers), LLaVA-1.5-13B (40 layers), Qwen2.5-VL-3B (36 layers), and Qwen2.5-VL-7B (28 layers).

Specifically, we extract the model’s internal attention to the original input image during inference using the method without incorporating visual thoughts (no-vis), and extract the internal attention to visual thoughts using methods that incorporate visual thoughts (n-lang, s-lang, e-img, g-img), following the approach illustrated in Figure 7(b) of the paper. This allows us to compare the shifting trend of attention after integrating visual thoughts.

We report the percentage of layers in which the attention score to visual thoughts exceeds that to the input image, as well as the percentage within the latter 50% of layers (shown in parentheses). The results are as follows:

Table 1: The percentage of layers in which the attention to visual thoughts exceeds that to the input original image. The number in parentheses indicates the result only discussed in the latter 50% layers of the model. || Layer |N-LANG(%)| S-LANG(%) | E-IMG(%)| G-IMG(%)| |:-:|:-:|:-:|:-:|:-:|:-:| | LLaVA-1.5-7B | 32 | 43.75 (12.50) | 40.63 (6.25) | 96.88 (100.00) | 93.75 (100.00) | | LLaVA-1.5-13B | 40 | 20.00 (0.00) | 35.00 (5.00) | 95.00 (100.00) | 92.50 (100.00) | | Qwen2.5-VL-3B | 36 | 97.22 (100.00) | 86.11 (100.00) | 72.22 (88.89) | 58.33 (100.00) | | Qwen2.5-VL-7B | 28 | 100.00 (100.00)| 92.86 (100.00) | 60.71 (100.00) | 60.71 (100.00) |

As shown in Table 1, for both the LLaVA and Qwen models with various number of layers, in majority of layers, the attention of visual thoughts exceeds that of the original image, demonstrating that the majority attention are distributed to visual thoughts during the reasoning process.

Furthermore, the results in "()" indicate that in the latter 50% layers of the model, attention to visual thoughts almost entirely surpasses that to the original image, even 100% visual thought attention is higher than original image, suggesting that visual thoughts can convey visual information more deeply than the original image itself. These findings are consistent with our observations in the paper.

* In the n-lang and s-lang settings for LLaVA, the model shows relatively low attention to visual thoughts due to the self-generated weaker language capabilities, resulting in captions and scene graphs that are overly simplistic and fail to accurately convey visual information.

We further compute the difference in attention score to visual thoughts relative to the original image, and average the results across all layers. The results are as follows:

Table 2: The difference in attention score to visual thoughts relative to the original image across all layers. || Layer | N-LANG | S-LANG| E-IMG | G-IMG| |:-:|:-:|:-:|:-:|:-:|:-:| | LLaVA-1.5-7B| 32 | -0.0529 | -0.0472 | 0.0156 | 0.0256| | LLaVA-1.5-13B | 40| -0.0481| -0.0275 | 0.0552 | 0.0637 | | Qwen2.5-VL-3B| 36| 0.1280| 0.0438| 0.0229 | 0.0388 | | Qwen2.5-VL-7B| 28 | 0.1069| 0.0553 | 0.0189 | 0.0152|

As shown in Table 2, both LLaVA (7B->13B) and Qwen(7B->3B) exhibit increased attention gains with deeper model layers. In the case of Qwen2.5-VL, the 3B variant, which has more layers, shows a more significant increase in attention compared to the 7B variant. This suggests that the impact of visual thoughts depends more on the specific model architecture than on the parameter scale.

Q2: The findings about information flow can be related to the LLAVA training pattern and it would be interesting to explore the different models.

R2: Thank you for your valuable comment. We follow your suggestion to further investigate the distribution of information flow under the integration of the four types of visual thoughts using Qwen2-VL-2B.

Table 3: Information flow distribution of Visual Thought and Image using Qwen2-VL-2B. || N-LANG| S-LANG| E-IMG| G-IMG| |-|:-:|:-:|:-:|:-:| | original image to reasoning | 0.0028 | 0.0007 | 0.0022 | 0.0022 | | visual thought to reasoning | 0.0056 | 0.0010 | 0.0038 | 0.0038 |

As shown in Table 3, we observe that visual thoughts can achieve larger information flow than the original image across all visual thought variants. Therefore, we can obtain the same conclusion as with the LLaVA models: “The information flow from visual thoughts to the reasoning stages is considerably stronger than the direct flow from the image to reasoning.”

Furthermore, we observe that, in contrast to the behavior of LLaVA, Qwen2-VL achieves higher information flow distribution in the n-lang setting, owing to its stronger language generation capabilities.

We will add more discussion in the next version.

Q3: Auto methods (e.g., CLIP) can be beneficial for conclusions within Figure 6.

R3: Thanks for your insightful feedback. To address the concerns regarding human evaluation bias and data scale limitations, we adopt your suggestion and use CLIP-based clarity scores to evaluate all visual thoughts.

Specifically, we compute the classification probability between two categories: “This is a clear (annotated logical) image” and “This is not a clear (annotated logical) image” (for G-IMG, content in parentheses is removed). These probabilities serve as quantitative indicators for clarity score of visual thought.

Table 4: Correlation analysis between clip-based clarity scores and model performance on GPT-4o-mini. || Pearson Correlation| P | |-|:-:|:-| | S-LANG | 0.692| 0.06 | | N-LANG | 0.693| 0.04 | | G-IMG | 0.731 | 0.04| | E-IMG | 0.646| 0.06|

As shown in Table 4, there is a clear correlation between clarity scores and model performance, indicating that high-quality visual thoughts are critical for achieving strong results. This aligns with our observations in paper.

Q4: The role of visual thoughts through representation analysis.

R4: Thank you for your constructive suggestions. Combining the results from Table 1 and Table 2, we observe that the internal attention mechanism reflects the effectiveness in conveying visual information of visual thoughts. For instance, in the case of LLaVA, which has relatively weaker language capabilities, the generated captions (N-LANG) or scene graphs (S-LANG) tend to be of lower quality and fail to effectively convey visual information, resulting in lower attention scores. This highlights the correlation between visual thoughts and the model’s internal representations.

Thanks for your acknowledgment and interest in our work! We sincerely appreciate your thorough and insightful comments on our work, and we will address each of your main concerns below:

Q1: The concept of visual thoughts is more an interpretive narrative than a substantive technical advance, like new methods and benchmarks.

R1: Thanks for your kind feedback. We sincerely believe there may be a misunderstanding. We sincerely think that the main contribution of our work is not to introduce new methods or benchmarks, but to present visual thoughts as a unifying perspective to better understand the mechanisms underlying MCoT (one of the most influential paradigms in the field). This contribution has also been acknowledged by other reviewers (Reviewer #NrVc, Reviewer #Z1x4, Reviewer #79f4). From this viewpoint, our study yields several notable and thought-provoking findings:

• First, we identify visual thoughts as the central functional component of MCoT. Regardless of their specific form, visual thoughts act as cache-like carriers of visual information during the reasoning process.
• Second, we conduct a comparative analysis of various MCoT implementations and reveal that the key determinants of performance are the clarity and efficiency of these visual thoughts.
• Third, we examine attention patterns and information flow to uncover the internal mechanisms by which visual thoughts serve as core intermediaries. Our findings suggest that they play a crucial role in sending more visual information to deeper model layers to produce the better reasoning trajectory of MCoT.

We hope this study can serve as a foundation for understanding how visual reasoning unfolds in large vision language models. We will definitely add more discussion in the next version to better address your technical concerns.

Q2: Benchmark directly against the most related prior works (e.g., Visual Sketchpad [15], Image-of-Thought [55], Scene Graph CoT) under comparable conditions.

R2: Thank you for your constructive feedback. Actually, the experiments shown in Figure 10 of our paper’s appendix were conducted directly on the Visual Sketchpad. In Table 1, our N-LANG setting is aligned with MM-CoT, the S-LANG setting corresponds to Scene Graph CoT, and the E-IMG setting is consistent with the Visual Sketchpad (note that the Image-of-Thought setting actually involves multi-turn visual reasoning).

In addition, to ensure what you referred to as “under comparable conditions,” we constrained all methods to generate responses within two interaction rounds, thereby ensuring comparability. Moreover, our comparison is not only direct but also accompanied by a detailed analysis of their differences and the underlying reasons.

Therefore, we have in fact conducted a comparison with the most relevant prior works under comparable conditions. We will revise the next version of the manuscript to clarify this point more explicitly.

Q3: Add an experiment showing that explicitly designing prompts or models informed by the “visual thoughts” perspective achieves measurable improvements over existing baselines.

R3: Thanks for your insightful feedback. Inspired by self-consistency, we follow your suggestion to design a novel visual-thought-guided strategy that ensembles diverse visual thoughts reasoning paths, which are refered as diverse-VT. As shown in Table 1, we observe that Diverse-VT achieves the best performance.

Table 1: Result of Divserse-VT. || MMVP | V*Bench-Position | V*Bench-attributes | M3CoT-physical | M3CoT-social | M3CoT-temporal | CoMT-deletion | CoMT-selection | CoMT-update | AVG | |:-|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| | w/o VT | 74.33 | 53.95 | 54.78 | 88.89 | 76.86 | 79.67 | 26.50 | 19.50 | 37.00 | 56.83 | | N-LANG (Maj@4) | 85.00 | 57.89 | 63.48 | 88.89 | 78.93 | 83.74 | 33.50 | 25.50 | 37.50 | 61.60 | | S-LANG (Maj@4) | 83.67 | 63.16 | 64.35 | 90.00 | 78.51 | 82.93 | 29.50 | 18.00 | 42.00 | 61.35 | | E-IMG (Maj@4) | 82.33 | 59.21| 65.22 | 92.22 | 78.10 | 86.18 | 33.50 | 28.00 | 49.00 | 63.75| | G-IMG (Maj@4) | 78.00 | 59.21 | 63.48 | 92.22 | 78.93 | 86.18 | 33.50 | 28.50 | 46.50 | 62.95| | Diverse-VT | 85.00 | 63.16 | 65.22 | 92.22 | 78.93 | 86.18 | 34.00 | 28.50 | 50.00 | 64.80 |

We will add more discussion in the next version.

Hope our clarifications can address your concerns, and we sincerely hope that you can reconsider our work in light of these clariìcations.