Thank you very much for the thoughtful and constructive review. We are encourged by the recognition of our paper’s clarity, comprehensive evaluation, and the value of our empirical study. We carefully address each of your concerns as below:

Q1: On the Superior Performance on Language Benchmarks

The strong language reasoning performance of OVR-7B can be attributed to the behavior-rich nature of the language data and the structure of our training pipeline:
1. Language Cold Start exposes the model to diverse and high-quality language reasoning traces rich in cognitive behaviors, laying a solid foundation for general reasoning.
2. Language RL serves as a cognitive scaffold, reinforcing these behaviors and aligning them exact-match reward signals.
3. Multimodal RL does not degrade language reasoning abilities. As shown in Lines 283–287, the model even slightly improves on language benchmarks after Stage 3, suggesting that multimodal adaptation complements, rather than interferes with, the reasoning skills acquired during earlier stages.

Q2: Extending Behavior Analysis on MathVerse

We have extended the same visual cognitive behavior analysis to MathVerse-MINI, following the methodology used in the main paper. The results show a consistent trend in the emergence and amplification of cognitive behaviors across training stages, supporting the generality of our findings across different tasks.

Q3: Ablation on Multimodal Reasoning Performance Across Stages

Table 5 presents the ablation results across training stages. To further address your concern, we have now extended the evaluation to include more benchmarks. These results further validate the contribution of each stage.

Q4: Qwen2.5-VL-7B Performance in Table 2

The performance of Qwen2.5-VL-7B is reported in Row 6 of Table 2. We have now added its evaluation on MMMU-Pro (38.3). OVR-7B achieves a 7.11 average improvement over the base model (44.45 --> 51.66) across all multimodal benchmarks in Table 2.

Q5: Difference with VL-Rethinker

While both VL-Rethinker and our method leverage RL to enhance reasoning in MLLMs, our OVR is distinct in motivation, analytical focus, performance, and methodology:

1. Cross-modal cognitive behavior transfer as the central goal: VL-Rethinker focuses on reflective fine-tuning within a vision-language setting to enhance multimodal reasoning. In contrast, our work uniquely centers on how cognitive behaviors acquired from language-only training can transfer to visual reasoning, revealing the internal mechanisms and generalization patterns behind advanced reasoning capabilies. This enables a scalable and interpretable pathway for building stronger multimodal reasoning systems.

2. In-depth behavioral analysis: We provide the first quantitative analysis of cognitive behavior emergence and cross-modal transfer, systematically tracking behavior shifts across stages. Notably, we discover that visual-specific cognitive behaviors can emerge in a zero-shot manner, even though they do not appear in the training data. This angle is not investigated in prior works.

3. Strong reasoning across modalities: Unlike prior approaches that focus solely on visual tasks, OVR also achieves SOTA performance across a suite of language reasoning benchmarks, demonstrating the value of modality-agnostic cognitive behaviors for strong reasoning capabilities.

Q6: Addressing Mentioned Weaknesses & Clarifying Contributions

We acknowledge that “RL with a cold start” pipeline is a well-established paradigm in LLM training. In our work, this pipeline is not positioned as the core novelty, but rather serves as a tool to systematically investigate the cross-modal transfer of cognitive behaviors and to build a strong multimodal reasoning model (lines 37–44).

We further clarify our distinct contributions as follows:

1. Cross-modal cognitive behavior transfer as the central focus: Our work uniquely centers on the unexplored behavior transfer from language to vision, revealing the internal mechanisms and generalization patterns behind advanced reasoning capabilies. This enables a scalable and interpretable pathway for building stronger multimodal reasoning systems.

2. In-depth behavioral analysis: We provide the first quantitative study of behavior emergence across training stages, including the surprising zero-shot emergence of visual-specific cognitive behaviors, transfer patterns in Sec. 6, and the correlation between behavior and reasoning performance (see Q3 response to Reviewer SEpa). This sheds light on how abstract behaviors generalize across modalities and benefit model's reasoning capabilities.

3. Superior reasoning across modalities: OVR achieves exceptional performance on both text and multimodal reasoning, as well as on general-purpose benchmarks, demonstrating the value of modality-agnostic cognitive behaviors for strong reasoning capabilities. Notably, it establishes a powerful and reproducible baseline for the open-source community, with full data, model and implementation details to be released publicly.

Thank you very much for the quick response and constructive suggestions. We have added additional experimental results to sincerely address each remaining concern:

• Q4: Comparison with Reproduced Results of the Baseline

  We fully agree with the rigorous and fair comparison and have replicated the results of Qwen2.5-VL-7B on all benchmarks under our evaluation framework. The table below reports our reproduced results and the performance of OVR-7B. OVR-7B achieves a significant +9.47 average improvement over the faithfully reproduced baseline, clearly demonstrating that the gains come from our proposed method rather than from differences in evaluation or optimization.

  	MathVista	MathVision	MathVerse	DynaMath	WeMath	LogicVista	MMMU-Pro	CharXiv-reas	CharXiv-desc	Avg
  Qwen2.5-VL-7B (Reproduce)	69.2	25.5	44.42	19.36	31.2	42.51	37.41	36.4	67.3	41.47
  OVR-7B	72.9	50.0	52.9	32.7	38.2	55.0	37.8	43.4	75.6	50.94

• Q5--Q6 Point3: Comprehensive Evaluation on Language Reasoning Benchmarks

  We appreciate your thoughtful suggestions and have conducted comprehensive evaluations of SOTA 7B-scale MLLMs on language reasoning benchmarks. As shown in the table, OVR-7B consistently outperforms 7B models across all tasks by a large margin, achieving the best results.

  	AIME 2024	AIME 2025	MATH500	GPQA Diamond	MMLU	MMLU-Pro
  Qwen2.5-VL-7B	6.7	6.7	67.4	31.8	69.6	51.7
  VLAA-Thinker-Qwen2.5-7B	7.5	1.25	58.4	33.84	67.98	49.96
  OpenVLThinker-7B	6.88	0.83	66.4	32.32	69.74	49.93
  MM-Eureka-Qwen-7B	5.63	1.88	67.80	39.39	72.46	51.65
  VL-Rethinker-7B	6.46	1.46	67.80	16.67	71.17	52.3
  OVR-7B	58.8	41.7	94.1	50.0	78.0	67.0

  This further strengthens the response (Point 3) to Q5–Q6. Different from prior works that primarily focus on visual tasks, OVR places equal emphasis on language reasoning, which we consider fundamental to developing robust and general intelligence, and a necessary foundation for effective multimodal reasoning. We also respect the contributions of concurrent efforts such as VL-Rethinker, which explore reasoning from other distinct perspectives.

We thank the reviewer again for the opportunity to further clarify and present our complete results, and apologize for the urgency of the previous response due to the experiment workload within the limited time. We sincerely hope these clarifications help strengthen your confidence in our submission. Please feel free to discuss if you have any other questions or suggestions.

(The following response provides an extended discussion beyond the original question, building upon the points outlined above.)

2. Does "more" cognition necessarily lead to better reasoning performance within the stage? (Further discussion inspired by Reviewer SEpa's Q3)

While the stage-wise results exhibit a clear positive correlation between behavior frequency and reasoning performance, we conduct a more rigorous intra-stage evaluation to faithfully address the nuance. Our analysis shows that within each training stage, behavior quantity does not directly correlate with performance gains.

This observation is described as follows:

• In cold start, visual re-inspection frequency rises sharply at first, then plateaus despite fluctuate performance improvements.
• In multimodal RL, behavior frequency drops initially, then rises again, ultimately surpassing prior levels. The reasoning performance increases with fluctuations.

While this initially appears counterintuitive, we provide further analysis to explain the underlying mechanisms:

(1) It’s not just quantity but also effectiveness that matters for behaviors

• Not all cognitive behaviors contribute equally to reasoning improvement. Some are effective, such as insightful re-evaluations that help correct mistakes; others are ineffective, like aimless repetitions.
• For instance, a model might generate statements like “Wait, maybe the answer is wrong…” and re-inspect the image, yet extract nothing useful. This distinction highlights that behavior effectiveness, rather than raw frequency, is the critical factor. This is also consistent with human cognitive intuition.

(2) Extended discussion: RL critically discerns and amplifies effective behaviors

• Cold start introduces a broad range of behaviors, transferring abstract reasoning patterns from language to vision. RL then acts as a filtering and amplifying mechanism, suppressing ineffective behaviors and selectively scaling up the most useful patterns (the crucial tokens analyzed in [2]).

---

To address this concern seriously, we will revise the paper to:

(1) Extend and clarify the analysis of the correlation between cognitive behaviors and reasoning ability, particularly in the Introduction and Section 6;

(2) Refine the wording regarding the role of cognitive behaviors in reasoning to avoid any vagueness or potential misunderstandings.

We thank the reviewer again for encouraging us to further clarify this point. Additionally, sometimes non-intuitive observations offer the greatest insights and deserve careful exploration. Thank you very much for your time and patience. Please feel free to share any further suggestions if you have.

We truly appreciate that our response helped resolve your concerns. Thank you so much for raising your confidence and we will definitely incorporate your valuable feedback into the revised paper to improve clarity.

Given the limited space and time during the initial rebuttal, some potential misunderstandings may remain regarding the relationship between cognitive behaviors and reasoning ability. As part of our commitment to a rigorous and faithful explanation, we offer a more detailed clarification here.

---

Please allow us to first clarify that

• Cognitive behavior transfer indeed contributes to stronger reasoning performance. Actually, Reviewer SEpa’s Q3 primarily focused on the role of behavior quantity. This is a valuable perspective, though it's equally important to consider behavior effectiveness.
• Our original response addressed this while also including several extended discussions (which, while not directly asked, we found valuable to share). We guess it's this mix that has caused some confusion.

To ensure clarity, we'd like to break down the original response for Reviewer SEpa Q3 more explicitly into two components:

• First, we confirmed the positive contribution of cognitive behavior transfer to stronger reasoning performance, as evidenced by clear stage-wise ablations. This explains the relationship between behavior transfer and the performance, and also provide a general response to the question focusing on the role of behavior quantity.

• Second (extended discussion for deeper insight), we offered a more nuanced discussion about:

  (1) whether "more" cognitive behaviors necessarily lead to better reasoning within individual training stages (a more precise exploration)

  (2) an extended analysis inspired by the question, focusing on intra-stage behavioral trends. While the latter is not directly asked, we believe it provides valuable perspective on the dynamics of behavior emergence and effectiveness during training.

Here is a brife and clear summary of the conclusions for your quick reference:

• (Main) Cognitive behavior transfer from language to vision is necessary to enable stronger multimodal reasoning performance, as demonstrated by the stage-wise improvements in ablations. In OVR, we facilitates this transfer and validates it as a practical and effective approach for achieving superior reasoning performance on both language and vision benchmarks.
• (Extension) However, a higher quantity of cognitive behaviors does not necessarily lead to better reasoning performance. Rather than focusing solely on behavior frequency, we emphasize the importance of behavior effectiveness—i.e., whether the behaviors meaningfully contribute to solving the task.

---

We elaborate the main conclusion in detail below:

1. Cognitive behavior transfer can indeed support stronger reasoning performance

Our results in Table 3 and Table 5 clearly demonstrate that cognitive behavior transfer contributes meaningfully to enhanced reasoning abilities:

• Zero-shot emergence: During the Cold Start stage, we observe a substantial improvement in visual reasoning performance (e.g., MathVision: 25.5 → 46.2; MME-R: 608.2 → 685). Simultaneously, visual-specific behaviors such as visual re-inspection emerge spontaneously in a zero-shot manner (from 0.0 to 2.0), indicating that structured reasoning patterns acquired in language can be transferred to the visual modality. This emergence, from absence to presence, highlights how behavior transfer directly facilitates improvements in visual reasoning.
• Progressive improvement across stages: When performance saturates at the current stage, continuing to the next stage training further improves both behavior frequency (0.0 → 2.0 → 2.4) and reasoning performance (Mathvision 25.5 → 46.2 → 47.5). This confirms the key point from our original response: visual cognitive behaviors are a necessary condition for achieving strong visual reasoning capabilities.

We are truly grateful to both Reviewer bA1c and Reviewer SEpa for your thoughtful engagement and academic rigor in discussing this relationship.

• We agree with Reviewer bA1c that our experiments primarily focus on correlation, not strict causation, and we sincerely apologize for any imprecise wording in our previous discussion. Following Reviewer SEpa's kind suggestion, we will be more careful with phrasing (e.g., replacing “necessary to enable” with “consistently associated with”) to avoid any potential misunderstanding.

• To clarify our revision for the paper:

  • The analyses in both the main paper and rebuttal are aimed to offer correlational insights into the relationship between cognitive behaviors and stronger multimodal reasoning—specially, which types of behavior emerge, when they emerge, and how they correlate with reasoning capabilities.
  • This line of analysis offers meaningful visual reasoning patterns to the community and may inspire further research into uncovering other key mechanisms underlying advanced MLLM reasoning.

We genuinely appreciate that reviewers' feedback helps deepen and strengthen the rigor of this important contribution.

---

Extended Lightweight Discussion on Causality

The reviewers’ feedback has helped us clarify that our analyses focus on correlation. Additionally, we genuinely regard Reviewer bA1c’s comment on causality as a valuable direction for future exploration. Below, we share a few ideas that could help investigate this in the future work:

This question can be precisely formulated as a causal inference problem: Is it (1) Behavior → Reasoning, or (2) Behavior ← Training → Reasoning?

To disentangle these possibilities, we find the following explorations promising:

• Intervention and Counterfactual Analysis: We can simulate a "do(B)" intervention by manually inserting or removing cognitive behaviors and evaluating whether this leads to a significant change in reasoning performance.
• Behavior Ablation Studies: By training models on de-behaviorized reasoning traces, we can compare them to models trained with full behaviors to assess the necessity of such behaviors for reasoning capabilities. A potential challenge here lies in controlling for linguistic consistency.
• Behavior Induction via Reward Shaping: Additionally, we can attempt to induce behaviors (e.g., reward injection) and measure whether they lead to stronger reasoning, which would further verify their causal role.

---

We sincerely thank Reviewer bA1c and SEpa again. We deeply respect your dedication and the time you’ve devoted to offering such valuable suggestions, which help make our work more comprehensive, in-depth, and rigorous.

This discussion became truly meaningful because of your valuable efforts.

Thank you very much for the constructive and encouraging feedback. We appreciate your recognition of the clarity of our writing, the significance of our cognitive behavior transfer findings, and the value of our analysis. We carefully address each of your concerns as follows.

Q1: Necessity of the Three-stage Training & Supporting Ablations

We conduct detailed ablations on RL training stages. As shown below and in Table 3 and Table 5, the results highlight the critical role of each component. Key observations include:

• Our two-stage RL training pipeline outperforms single-stage or mixed-stage RL settings, achieving the best average results on both language and multimodal reasoning tasks.
• Stage 3 further enhances both language and multimodal reasoning, while also promoting the emergence of more visual cognitive behaviors as shown in Table 3.

1. The experiments results:

• Our proposed training pipeline, incorporating both Stage 2 and Stage 3, outperforms the mixed RL setting, yielding improvements on vision reasoning tasks such as MathVision (+0.5) and MathVerse (+1.0).
• Each stage contributes positively across multiple reasoning benchmarks, validating the necessity of our training approach.
• Performing Stage 3 after Stage 2 not only enhances visual reasoning and elicits more visual cognitive behaviors (as shown in Table 3), but also further boosts language reasoning, with gains on AIME2024 (+0.9) and MATH500 (+0.5).

2. Discussion on Two-stage RL vs Mixed RL:

• Beyond performance, separating the stages is also more practical from a training efficiency perspective: language RL involves longer and more complex responses, requiring longer training time; multimodal RL, while not trivial, requires shorter response length and exhibits less scaling complexity (as discussed in line 298-303), resulting in faster convergence. Notably, recent works such as NemoTron-1.1 [1] have adopted similar staged training strategies, achieving promising results.
• Discussion: While mixed RL is a valid alternative, it introduces greater training cost without clear advantages. Further scaling with prolonged RL has the potential to amplify the differences between mixed and separate RL stages and lead to stronger results, though it would require significantly more compute resources.

[1] AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy

Q2: Comparison to RL-Based Methods

We complement evaluations for representative RL-based baselines in the table. Our method achieves average gains of 6.82 on vision reasoning tasks** over these RL-based baselines and 25.95 on text reasoning tasks over the base model, demonstrating its effectiveness across modalities.

Q3: Relationship between Behaviors and Reasoning Performance

We analyze both stage-wise and intra-stage correlations between cognitive behavior frequency and reasoning performance to address this question. The key conclusions includes:

• Cognitive behavioris a necessary but not sufficient condition for strong reasoning abilities.
• Higher behavior frequency does not necessarily imply better reasoning, as performance also depends on the task-alignment and effectiveness of the emerging behaviors.

The detailed observations and analysis are as below:

1. Strong stage-wise correlation

• Observation: Across training stages, both the visual re-inspection behavior and reasoning performance increase progressively. The behavior frequency grows from 0.0 → 2.0 → 2.4, while MathVision performance improves from 25.5 → 46.2 → 47.5.
• Analysis: This stage-wise trend suggests a strong positive correlation between cognitive behavior frequency and reasoning capabilities. However, correlation does not imply direct causation. We futher analyze intra-stage performances as follows.

2. Intra-stage divergence highlights behavior effectiveness

• Observation:
  • During cold start, behavior frequency increases rapidly at first and then plateaus with fluctuations (e.g., visual re-inspection behavior: 0.3 → 2.2 → 2.0).
  • In Multimodal RL, we observe an initial drop in behavior frequency, followed by a gradual recovery (2.4 → 0.5 → 2.5), yet reasoning performance continues to climb throughout.
• Analysis: This divergence reveals that a greater number of behaviors do not necessarily lead to better reasoning. In early Cold Start, the model generates a wide range of exploratory behaviors, regardless of whether they are redundant or effective. As training progresses, we observe that behavior frequency plateaus in late cold start and drops at the start of RL, yet performance continues to improve. This indicates that gains are driven more by behavior effectiveness than by behavior quantity.
• Further discussion: In the RL stage, the training initially reshapes the model's behaviors to maximize rewards and suppresses ineffective reasoning paths, leading to a temporary drop in behavior frequency. However, the recovery shows RL critically discerns and scales up effective patterns (the crucial tokens analyzed in [2]). This aligns with the insight that SFT memorizes while RL generalizes [3].

We will add the correlation analysis in the revision.

[2] Reasoning with exploration: An entropy perspective
[3] SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training

We sincerely appreciate the reviewer’s constructive suggestions for improving the clarity and precision of our writing, and we are encouraged that our previous response helped address your first two concerns

Given the limited space and time during the initial rebuttal, some potential misunderstandings may remain regarding the relationship between cognitive behaviors and reasoning ability. As part of our commitment to a rigorous and faithful explanation, we would like to offer a more detailed clarification here.

---

Please allow us to first clarify that:

• Cognitive behavior transfer indeed contributes to stronger reasoning performance. However, Q3 primarily focused on the role of behavior quantity. This is a valuable perspective, though it's equally important to consider behavior effectiveness.
• Our original response addressed this while also including several extended discussions (which, while not directly asked, we found valuable to share). We guess it's this mix that has caused some confusion.

To ensure clarity, we explain that our original response is actually structured around two key points:

• First, we confirmed the positive contribution of cognitive behavior transfer to stronger reasoning performance, providing a general response to Q3 regarding the role of behavior quantity.

• Second, we offered a more nuanced extended discussion:

  (1) addressing whether "more" cognitive behaviors necessarily lead to better reasoning within individual training stages (a more precise extended exploration to Q3)

  (2) a further analysis inspired by Q3, focusing on intra-stage behavioral trends. While the latter is not directly asked in Q3, we believe it provides valuable perspective on the dynamics of behavior emergence and effectiveness during training.

Here is a brife and clear summary of the conclusions for your quick reference:

• (Main) Cognitive behavior transfer from language to vision is necessary to enable stronger multimodal reasoning performance, as demonstrated by the stage-wise improvements in ablations. In OVR, we facilitates this transfer and validates it as a practical and effective approach for achieving superior reasoning performance on both language and vision benchmarks.
• (Extension) However, a higher quantity of cognitive behaviors does not necessarily lead to better reasoning performance. Rather than focusing solely on behavior frequency, we emphasize the importance of behavior effectiveness—i.e., whether the behaviors meaningfully contribute to solving the task.

---

We elaborate in detail for the main conclusion below :

1. Cognitive behavior transfer can indeed support stronger reasoning performance

Our results in Table 3 and Table 5 clearly demonstrate that cognitive behavior transfer contributes meaningfully to enhanced reasoning abilities:

• Zero-shot emergence: During the Cold Start stage, we observe a substantial improvement in visual reasoning performance (e.g., MathVision: 25.5 → 46.2; MME-R: 608.2 → 685). Simultaneously, visual-specific behaviors such as visual re-inspection emerge spontaneously in a zero-shot manner (from 0.0 to 2.0), indicating that structured reasoning patterns acquired in language can be transferred to the visual modality. This emergence, from absence to presence, highlights how behavior transfer directly facilitates improvements in visual reasoning.
• Progressive improvement across stages: When performance saturates at the current stage, continuing to the next stage training further improves both behavior frequency (0.0 → 2.0 → 2.4) and reasoning performance (Mathvision 25.5 → 46.2 → 47.5). This confirms the key point from our Q3 response: visual cognitive behaviors are a necessary condition for achieving strong visual reasoning capabilities.

(The following response provides an extended discussion related to this question, building upon the points outlined above.)

2. Does "more" cognition necessarily lead to better reasoning performance within the stage? (Further discussion inspired by Q3)

While the stage-wise results exhibit a clear positive correlation between behavior frequency and reasoning performance, we conduct a more rigorous intra-stage evaluation to faithfully address the nuance. Our analysis shows that within each training stage, behavior quantity does not directly correlate with performance gains.

This observation is described as follows:

• In cold start, visual re-inspection frequency rises sharply at first, then plateaus despite fluctuate performance improvements.
• In multimodal RL, behavior frequency drops initially, then rises again, ultimately surpassing prior levels. The reasoning performance increases with fluctuations.

While this initially appears counterintuitive, we provide further analysis to explain the underlying mechanisms:

(1) It’s not just quantity but also effectiveness that matters for behaviors

• Not all cognitive behaviors contribute equally to reasoning improvement. Some are effective, such as insightful re-evaluations that help correct mistakes; others are ineffective, like aimless repetitions.
• For instance, a model might generate statements like “Wait, maybe the answer is wrong…” and re-inspect the image, yet extract nothing useful. This distinction highlights that behavior effectiveness, rather than raw frequency, is the critical factor. This is also consistent with human cognitive intuition.

(2) Extended discussion: RL critically discerns and amplifies effective behaviors

• Cold start introduces a broad range of behaviors, transferring abstract reasoning patterns from language to vision. RL then acts as a filtering and amplifying mechanism, suppressing ineffective behaviors and selectively scaling up the most useful patterns (the crucial tokens analyzed in [2]).

---

To address your concern seriously, we will revise the paper to:

(1) Extend and clarify the analysis of the correlation between cognitive behaviors and reasoning ability, particularly in the Introduction and Section 6;

(2) Refine the wording regarding the role of cognitive behaviors in reasoning to avoid any vagueness or potential misunderstandings.

We thank the reviewer again for encouraging us to refine this point more precisely. Additionally, sometimes non-intuitive observations offer the greatest insights and deserve careful exploration. We sincerely hope that these clarifications help strengthen your confidence for the acceptance of our work. Please feel free to share any further suggestions.

We thank the reviewer for the constructive feedback and recognition of our paper’s strong performance and interesting research angle. We carefully address each of your concerns below.

Q1: More Ablations

We conducted comprehensive ablations along three dimensions: (1) different cold-start data strategies, (2) RL training schedule, and (3) enhanced cold start with self-distilled multimodal data.

The results in the table are summarized:

• Adding open-source multimodal cold-start data or applying mixed RL stages does not outperform our proposed training pipeline.
• Introducing self-distilled data from OVR into cold start leads to improved performance.

The ablation details are shown below:

1. Ablations on Cold Start (Part I)

• Settings: We experiment with training on multimodal cold-start data either independently or in combination with language cold-start data. The multimodal cold-start datasets contain publicly available sources such as R1-OneVision [1] and Vision-R1 [2], which are constructed by either synthesizing image-text pairs using image descriptions and language reasoning models or rendering language cold-start data into multimodal form.
• Results & Analysis:
  • Training with multimodal cold-start data alone leads to lower performance on both language and multimodal reasoning tasks, particularly in language reasoning (e.g., AIME2024: 23.1 vs. 54.4). This reveals the lack of high-quality multimodal cold-start datasets. The much better performance on MathVision also verifies that language reasoning skills are fundemental for advanced multimodal reasoning, and demonstrates the robustness of our data curation pipeline.
  • Adding multimodal cold-start data on top of language cold-start training reduces performance. This arises from conflicting reasoning patterns: multimodal cold-start data has obvious differences in style, length and structure compared to the language-only version. Nevertheless, this challenge can be addressed through large-scale RL and iterative self-distillation.
• Discussion: Constructing effective multimodal cold-start data remains a significant challenge. Unlike the language domain, the multimodal community lacks high-quality, open-source reasoning teachers like DeepSeek-R1, and most closed-source models hide explicit reasoning traces. Our method strategically avoids this limitation by leveraging a strong language cold-start phase followed by multimodal RL, resulting in a robust and scalable multimodal reasoner.

2. Ablations for RL stages (Part II)

• Settings: Based on the language cold start , we ablate three RL strategies: (1) language-only RL, (2) multimodal-only RL, and (3) a one-stage mixed RL that combines both.
• Results & Reasons: The two-stage RL setup consistently yields the best performance across reasoning benchmarks. From a training efficiency perspective, separating the stages is also more practical: language RL involves longer and more complex responses, requiring longer training time; multimodal RL, while not trivial, requires shorter response length and exhibits less scaling complexity (as discussed in line 298-303), resulting in faster convergence. Notably, recent works such as NemoTron-1.1 [3] have adopted similar staged training strategies, achieving promising results.
• Discussion: While mixed RL is a valid alternative, it introduces greater training cost without clear advantages. Further scaling with prolonged RL has the potential to amplify the differences between mixed and separate RL stages and leads to stronger results, though it would require significantly more computation resources.

3. Enhancing Cold Start with Self-Distilled Multimodal Data (Part III)

• Setting: To address the reviewer’s suggestion on “(5) multimodal cold start with more data,” we augment the cold start with self-distilled multimodal data generated by our own OVR model. As discussed in Point 1, the community currently lacks high-quality multimodal cold-start data and is challenged in constructing. Thus, we leverage the model’s own capabilities to distill high-quality reasoning traces, and train with a combination of language-only cold-start data and self-distilled multimodal samples.
• Reults & Analysis: Incorporating self-distilled data during cold start leads to better performance in both language and multimodal reasoning tasks, faster convergence, and lower training loss. These demonstrate the effectiveness of self-distillation [4] in bootstrapping multimodal reasoning. Moreover, they point to a promising direction: in the absence of strong open-source multimodal teacher models, iterative RL and self-distillation can serve as a viable and scalable strategy for progressively enhancing model capabilities.

4. Cognitive Behaviors on ablations: Given that behaviors primarily emerge during cold start, we analyze their development across the three cold start variants. Consistent with final performance metrics, the multimodal cold start does not yield an increase in either textual or visual cognitive behaviors. We attribute this to the following factors:

• Multimodal cold-start data still suffers from limited quality. The synthetic nature fails to introduce new reasoning patterns, which even hurts performances.
• Visual behaviors tend to emerge in the early steps of cold start and plateau thereafter (visual re-inspection: 0.3 → 2.2 → 2.0 from iter 0 → iter 200 → iter 700 in OVR). Since multimodal cold start data is derived from language-only sources, the underlying patterns remain unchanged. As a result, simply adding more data, regardless of modality, does not further enhance the emergence of cognitive behaviors.

[1] R1-Onevision: Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization
[2] Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models
[3] AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy
[4] DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

Q2: Cause of Transfer after Cold Start

Our analysis suggests that the transfer is primarily driven by the nature of cold start data, rather than data volume.

1. Transfer originates from linguistic “mental imagery”: The linguistic data distilled from DeepSeek-R1 often include internal visualizations (somehow reveals its mental imagery) to support mathematical reasoning, which are often articulated through phrases such as “let me visualize/sketch...”. Once this linguistic scaffolding was introduced into MLLM, these patterns were rapidly grounded in actual visual input for handling multimodal reasoning tasks.

2. Emergence happens early and is not driven by data size: We track the emergence of visual behaviors across different iterations within the cold start. Specifically, we observe that a significant portion of visual behavior occurs at the early step and then fluctuates rather than steadily increasing (0.3 → 2.2 → 2.0 from iter 0 → iter 200 → iter 700). This suggests that the effect is not proportional to data size, but instead reflects early inductive bias amplification.

Q3: More Perception-centric Evaluations

We evaluate on more perception benchmarks. OVR achieves consistent improvements than the baseline in the table.

Q4: Generalization across Models and Data Sizes

Our extended experiments support the core finding—cognitive behaviors learned in language can be transferred to vision for advanced visual reasoning—generalizes well across both larger model scales and more data with broader task coverage.

1. Model Scaling: Our training on an in-house MLLM built upon Qwen2.5-32B confirms strong scalability, boosting AIME 2024 performance from 13 → 73 and MathVision from 35 → 60.

2. Task & Data Scaling: Rather than simply increasing data volume, we also emphasized task diversity and data quality. By adding just ~10k high-purity STEM samples to the RL training, we achieve a +2% increase on MathVision.

These results indicate that the benefits of behavior transfer not only generalize but also amplify with larger model capacity and curated data, reinforcing the robustness and scalability of our approach.

Q5: Reward Design

Using exact match as the reward function is well-suited to the nature of the studied tasks and offers robustness to reward hacking.

First, the tasks we focus on includes multimodal math, logic, and VQA tasks, which typically have objectively correct answers. Exact match provides the most direct and reliable correctness signal.

Second, from a broader methodological perspective, exact match is a widely adopted reward shaping in reasoning tasks. It offers an unambiguous reward signal, thereby effectively mitigating reward hacking. This choice is consistent with mainstream methods such as Open-Reasoner-Zero [5].

While we agree that reward design is an important area for future exploration, it is not the focus of this work. In future studies, we plan to investigate hybrid reward strategies that integrate model-based feedback and rule-based signals to further enhance robustness.

[5] Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model

Dear Reviewer 9Ror,

We sincerely appreciate the time and great efforts you have dedicated to reviewing our work.

If possible, we would be grateful for the opportunity to further engage in discussion and hear your thoughts. We would love to know whether our experiments and explanations have addressed your concerns, or if you have any additional suggestions for improvement. Thank you!

Best,

The Authors

Thank you for the detailed and insightful feedback. We appreciate the reviewer's recognition of our central hypothesis as "compelling", our methodology as "logical and well-structured", our results as "impressive", and our behavioral analysis as providing "valuable insight". Below, we carefully address each of your concerns and clarify potential misunderstandings.

Q1: Evaluation Reliability

We conducted a human evaluation to validate the reliability of GPT-4o-mini as the behavior classifier, focusing on visual re-inspection.

• The results show high alignment between the model and human predictions.
• Moreover, the human annotations further support the effectiveness of our behavior transfer.

1. Setup: We randomly sampled 100 queries from the MathVision-mini benchmark and collected model outputs from both Qwen2.5-VL-7B and OVR-7B (200 responses in total). These anonymized samples (image, query, response) were shuffled and evaluated on a dedicated annotation platform by five professional annotators with expertise in CV and NLP. Annotators identified whether a given response exhibited visual re-inspection behavior, based on formal definitions and detailed annotation criteria (formatted as prompts in Appendix S-Figure 1).

2. Metrics:

• We first compute the behavior classification accuray of GPT-4o-mini against aggregated human annotations.
• Additionally, we compared the behavior frequencies of the two models, using GPT and human annotations respectively.

3. Results:

• GPT Classification Accuracy: Across all annotated cases, GPT-4o-mini achieved an accuracy of 92.42%, showing strong alignment with human judgments. The model produced only 2 false positives and 13 false negatives, showing a conservative tendency in behavior detection. This indicates that GPT-4o-mini reliably captures the overall trend of behavior emergence, supporting its use as a practical and scalable proxy for behavior classification in our study.

• Behavior Frequency Comparison: The table shows the behavior frequency of the two models, as measured by GPT and human annotations respectively. Both sources consistently indicate that OVR demonstrates a substantially higher behavior frequency than the baseline, further validating that the behavior was effectively transferred. | Model | GPT (%) | Human (%) | |----------------|---------|-----------| | Qwen2.5-VL-7B | 0.0 | 2.0 | | OVR-7B | 10.1 | 19.2 |

Q2: Distinguishing Genuine Behavior from Pattern Mimicry

We appreciate the opportunity to address the possible misunderstanding. We emphasize that the defined visual-specific cognitive behaviors has been evaluated in the main paper to handle this concern, which are not present in the training data. The analysis on emergence of these behaviors demonstrates the genuinely learned cognitive process, rather than superficial mimicry. We further clarify the definition and findings as follows:

1. Visual-specific behaviors defined to capture genuine transfer: To assess whether the model genuinely learns and transfers cognitive behaviors, we examine visual-specific behaviors grounded in visual inputs (defined in Sec. 5.4 and detailed in Appendix C.1). While these behaviors are conceptually related to their linguistic counterparts (explained in S-Table 1), they are fundamentally different in modality, making them well-suited for verifying true cross-modal transfer rather than superficial mimicry.

2. Zero-shot emergence confirms internalization and transfer: These behaviors are entirely absent from all training data. Even During the multimodal RL, the model only receives ground truth answers without any intermediate reasoning steps or behavioral scaffolding. Nevertheless, as shown quantitatively in Table 3 and qualitatively in Figure 2, these behaviors consistently emerge. This strongly suggests that the model has internalized and strategically transferred abstract reasoning strategies to the visual domain, rather than mimicking surface-level text artifacts. We consider this one of the most valuable findings of our work.

We will further clarify this in the revision to avoid any ambiguity.

Q3: Contribution Clarification

We acknowledge that “RL with a cold start” is a well-established paradigm in LLM training. In our work, this pipeline is not positioned as the core novelty, but rather serves as a tool to systematically investigate the cross-modal transfer of cognitive behaviors and to build a strong multimodal reasoning model (lines 37–44).

We further clarify our distinct contributions as follows:

1. Cross-modal cognitive behavior transfer as the central focus: Our work uniquely centers on the unexplored behavior transfer from language to vision, revealing the internal mechanisms and generalization patterns behind advanced reasoning capabilies. This enables a scalable and interpretable pathway for building stronger multimodal reasoning systems.
2. In-depth behavioral analysis: We provide the first quantitative study of behavior emergence across training stages, including the surprising zero-shot emergence of visual-specific cognitive behaviors (see Q2 response), transfer patterns (Sec. 6), and the correlation between behavior and performance (see Q3 response to Reviewer SEpa). This sheds light on how abstract behaviors generalize across modalities and benefit model's reasoning capabilities.
3. Superior reasoning across modalities: OVR achieves exceptional performance on both text and multimodal reasoning, as well as on general-purpose benchmarks, demonstrating the value of modality-agnostic cognitive behaviors for strong reasoning capabilities. Notably, it establishes a powerful and reproducible baseline for the open-source community, with full data, model and implementation details to be released publicly.

Q4: On the Reasoning-Perception Trade-off

We would like to emphasize that the primary goal of this paper is to enhance multimodal reasoning. Within this context, the partial decline in perception metrics is acceptable and solvable, and therefore should not be considered a “major flaw”. We address this concern through four key aspects:

1. No "consistent degradation" in perception

• Contrary to the reviewer's comments, MMVet scores in Table 5 actually increase from 57.3 to 60.8.
• We also evaluate on more perception-centric benchmarks, and the results show that OVR maintains or even improves performance across these tasks. | Model | BLINK | HallusionBench | MMbench | |------------------|--------|----------------|---------| | Qwen2.5-VL-7B | 53.71 | 49.43 | 85.84 | | OVR-7B | 54.13 | 54.90 | 86.42 |

2. Acceptable trade-off for the reasoning-focused OVR: Given our framework’s emphasis on language and multimodal reasoning, some marginal drop on perception metrics is an acceptable trade-off. While MME-P shows a decline of only 7.3%, reasoning metrics improve more substantially. For instance, AIME24 increases from 6.67 → 58.8 and MathVision from 25.5 → 50.0, demonstrating a strong overall capability gain.

3. Solution through Scaling: As discussed in line 280-282, we further scale up multimodal RL, leading to a gradual recovery in perception metric MME-P 1547→1600 and a increase in reasoning metrics like MME-R 713.6 → 725. This suggests that perception degradation is not fundemental, but solvable through continued training.

4. A well-studied and explainable phenomenon in the community

• Perception-reasoning trade-off is a well-recognized phenomenon in MLLM reasoning [1]. The observed decline in perception performance often arises from reduced reliance on fine-grained visual details, as models develop longer and more abstract reasoning chains that favor linguistic priors.
• We plan to address this in future work via prolonged multimodal RL and self-distillation, aiming to jointly reinforce reasoning and perception abilities.

[1] More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models

Q4

To avoid any potential misunderstanding, we’d like to clarify that the primary focus of our work is on enhancing MLLM reasoning from the angle of cognitive behavior transfer. We do consider perception an important capability—which is why we faithfully include it as a side discussion in the paper—but it's actually not the central focus. Beyond clarification, we add further discussion on this extra part in the point 3 below to address reviewer's concerns.

The concrete explanations are detailed below:

1. As a paper focused on reasoning, our work presents a clear and distinctive motivation, strong experimental evidence, and in-depth analytical insights.

• Built around the novel perspective of cognitive behavior transfer, we propose a distinct training principle, conduct extensive evaluations, and provide in-depth behavioral analysis. We sincerely appreciate that all reviewers have acknowledged the coherence and value of these efforts for the mainline.

2. The perception-related discussion is an additional point but not the mainline.

• This part in the main paper was actually included to faithfully report on an observed phenomenon. It is not central to the main objective of improving reasoning in MLLMs via cognitive behavior transfer and is discussed to sincerely motivate future community efforts.

3. Further discussion, analysis, and solutions on this point are as follows (though not our main focus):

While this is not our central focus, we agree with the reviewer that the "trade-off" represents an important direction for future research. Therefore, in our original response to Q4, we conducted additional study and observed the following:

• This phenomenon has been well-studied and analyzed in [4,5] for current SOTA visual reasoning models [6,7], which exhibit perception degradation. This is a feature of the problem space, not a "flaw" in our method.
• There is actually no fundamental degradation in perception. Conversely, we observe improvements on perception-related benchmarks such as BLINK and HallusionBench.
• The practical solution proposed in Point 3 of our original response has been shown effective in addressing the temporary and marginal drop observed solely on MME-P, demonstrating that this issue is addressable and won't be a future concern.

[4] More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models

[5] Are Reasoning Models More Prone to Hallucination?

[6] MM-Eureka: Exploring the Frontiers of Multimodal Reasoning with Rule-based Reinforcement Learning

[7] OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles

We sincerely thank the reviewer again for your time and thoughtful engagement throughout the rebuttal process. We truly hope these clarifications help resolve any remaining ambiguities or misunderstandings. Please feel free to reach out if you have any further concerns or suggestions.

Extended cold start experiments for supporting OVR's distinct and impactful training principle:

(Supporting Point 1 in the Part I above)

1. Settings: We experiment with training on multimodal cold-start data either independently or in combination with language cold-start data. The multimodal cold-start datasets contain publicly available sources such as R1-OneVision [2] and Vision-R1 [3].

2. Results: Training with multimodal cold-start data alone leads to lower performance on both language and multimodal reasoning tasks. Adding multimodal cold-start data on top of language cold-start training reduces performance. This reveals the lack of high-quality multimodal cold-start datasets.

3. Conclusion: Constructing effective multimodal cold-start data remains a significant challenge. Unlike the language domain, the multimodal community lacks high-quality, open-source reasoning teachers like DeepSeek-R1, and most closed-source models hide explicit reasoning traces.

[2] R1-Onevision: Advancing Generalized Multimodal Reasoning through Cross-Modal Formalization

[3] Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

We sincerely appreciate that our response helped address your first two concerns. We also thank the reviewer for the opportunity to further clarify our contributions, and avoid any vagueness or potential misunderstanding caused by the limited space in the initial rebuttal.

Q3

To help better understand the core contributions of our work, we would like to offer a clearer clarification of our original response to Q3. The following elaboration is not a repetition, but a faithful extended clarification.

Our paper's contribution is multifaceted, extending beyond novel perspectives to include new conceptual paradigms, new scientific findings, and powerful practical artifacts:

1. A New Conceptual Paradigm: Language-to-vision cognitive transfer, the central focus of our work, is indeed a distinct and impactful training principle for advanced MLLM reasoning

• We totally understand your concern that “RL with a cold start” is often regarded as a common tool in LLMs. However, its direct application to MLLMs fails to yield comparable benefits, primarily due to: (clearly demonstrated by the cold start experiments in the next box)
  • The scarcity of high-quality multimodal cold-start data
  • Unlike the language domain, the multimodal community lacks powerful open-source reasoning teachers like DeepSeek-R1, and most closed-source models hide explicit reasoning traces.
• Our cognitive transfer paradigm directly addresses this critical gap from a novel and underexplored perspective. By first cultivating effective reasoning patterns in the data-rich language domain, It provides an effective strategy to leverage existing, high-quality resources to enhance reasoning capabilities across both modalities.

In summary, our work demonstrates a scalable, interpretable, and reproducible training paradigm to build stronger multimodal reasoners. This distinct principle will meaningfully impact the future direction of MLLM reasoning. This is a fundamental shift in how to approach the problem.

2. New Scientific Insights: In-depth behavior transfer analysis reveals patterns essential to reasoning

• As shown in the Q3 response (point 2), we present the first quantitative study on cross-modal behavior transfer, tracking how cognitive behaviors learned in language emerge and evolve in the visual domain. This is a direct and novel contribution to the community's scientific knowledge.
• Beyond validating the transfer itself, our analysis reveals key reasoning-effective patterns within long multimodal reasoning chains. This acutually offer practical principle for constructing stronger multimodal cold-start data, highlighting which behavior patterns are worth preserving or distilling. This also helps address the data scarcity challenge discussed in the first point.

3. A Powerful Community Resource: We developed a strong open-source reasoner for both language and vision

• As shown in the Q3 response (point 3), unlike prior works that primarily focus on visual reasoning, our OVR demonstrates superior reasoning capabilities across both language and vision tasks.
• The resulting OVR is valuable not only from the perspective of cognitive behavior transfer, but also due to the effective data curation pipeline and our robust RL training methodology. Creating a powerful, accessible, and competitive baseline is a significant form of novelty that will directly faciliate future research.

Finally, regarding whether a paper offers sufficient contribution, we borrow a perspective from Novelty in Science [1]:

If you hear a good idea, there is a moment of surprise and then, the better it is, the more obvious it may seem. If it is easy to explain and obvious in hindsight, this in no way diminishes the creativity (and novelty) of the idea.

[1] Black. Novelty in Science. Medium. https://medium.com/@black_51980/novelty-in-science-8f1fd1a0a143