Praxis-VLM: Vision-Grounded Decision Making via Text-Driven Reinforcement Learning

We sincerely thank you for the valuable feedback and for recognizing the strengths of our work, including the interesting empirical discovery, strong performance, and thorough analysis. Below we address the raised questions and concerns.

Q1: Not Directly Optimizing Visual Grounding

This is a good question! We would like to first clarify that our approach focuses on transferring reasoning skills learned from text-only training to vision-grounded multimodal inference. Our preliminary analysis (Section 2) shows that reasoning and decision-making can be disentangled from visual perception, allowing us to leverage the scalable text-only data to enhance VLM decision-making abilities.

This design addresses a critical bottleneck in current VLM development: the lack of accessible, high-quality multimodal data for model training, especially for decision-making tasks. By focusing on text-driven training, our method allows for efficient and cost-effective learning of reasoning skills without relying on multimodal supervision.

As such, directly optimizing visual grounding during training falls outside the scope of this work, as it would require a fundamentally different setup involving paired multimodal data, which contrasts with our case where such multimodal training data is unavailable.

As acknowledged in our Limitation Section (Line 550), we agree that further strengthening visual perception is an important next step for improving vision-grounded decision-making. While beyond the scope of this work, we plan to explore perception-focused training in future research.

---

Q2: Reliance on a Strong Backbone VLM like Qwen2.5-VL

This is a good question and we would like to clarify our choice of Qwen2.5-VL.

First, due to computational constraints, we focus on models at the 3B and 7B scales. Qwen2.5-VL is selected because it is a strong, open-source VLM with competitive performance, offering a reproducible and reliable foundation for our study. Recent work [1,2] suggests that RL mainly elicits and refines knowledge embedded in the base model, rather than learning entirely new capabilities. Therefore, tackling complex decision-making depends critically on the strength of the base model, and Qwen2.5-VL provided the best open-source option at this scale when our work is conducted.

Second, we follow established practice in reasoning VLM research, including OpenVLThinker [3], R1-OneVision [4], and MM-Eureka [5], all of which also build upon Qwen2.5-VL. This ensures comparability and consistency with existing methods.

We agree that extending Praxis-VLM to other architectures is an important future direction. However, our primary goal here is to demonstrate the principle and effectiveness of text-driven reinforcement learning for enhancing reasoning and decision making. We will clarify this motivation and model choice more explicitly in the revision.

[1] Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?

[2] S1: Simple test-time scaling

[3] OpenVLThinker: An Early Exploration to Complex Vision-Language Reasoning via Iterative Self-Improvement

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

[5] MM-Eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning

---

L1: Regarding Data Generation

Scalable data generation pipeline

This is a good point! While our data generation relies on an off-the-shelf model (GPT-4o), the pipeline is intentionally designed to be both diverse and scalable. We use an automated process: GPT-4o is prompted to generate 10 samples per batch, followed by deduplication based on word-overlap similarity to remove highly similar scenario descriptions within each batch. Beyond this step, we avoid manual filtering or heavy curation, enabling fast, domain-agnostic dataset creation that can scale to large volumes.

Leveraging LLMs for synthetic generation has been widely adopted in previous research [1]. Our approach builds on this by further eliminating the need for costly image-text paired datasets, offering a more practical and data-efficient strategy to enhance VLM reasoning capabilities.

Diversity of synthetic scenarios

To assess the diversity of the generated scenarios, we prompt GPT-4o to cluster the generated textual situations into topical categories. The results indicate a broad coverage across varied scenarios:

• Workplace Performance and Personal Issues,
• Resource Allocation
• Project Management
• Balancing Competing Interests
• Policy, Rules & Enforcement
• Ethical Dilemmas
• Interpersonal Conflict
• Emergency Handling
• Navigating Setbacks
• Event Planning & Logistics
• Balancing Inclusivity and Majority Preference

Impact of Synthetic Data on Training

While the generated data cover diverse situations, there can still be a domain gap when compared to specific evaluation benchmarks (e.g., PCA-Bench’s embodied robotics or EgoNormia’s egocentric video contexts). However, Praxis-VLM demonstrates consistent improvements and robust generalization across all benchmarks, indicating that our text-driven RL training learns generalized, transferable reasoning skills rather than overfitting to domain-specific patterns. This is also supported by the reasoning dimension analysis in Figure 5, showing that even without deliberate domain-specific data curation, Praxis-VLM successfully acquires adaptable reasoning abilities.

As discussed in the Limitation Section (Line 545), we believe exploring advanced data generation and selection methods for more efficient model training is an important direction. We also appreciate your insightful suggestion of how potential GPT hallucinations affect training, which we plan to investigate in future work. We will revise the Limitation Section to further incorporate your suggestions into our discussions.

[1] Self-Instruct: Aligning Language Models with Self-Generated Instructions

---

L2: Regarding Established Approach and focus more on the engineering perspective

Thank you for raising this point. While we acknowledge that components such as GRPO, and multi-stage training are established techniques, we respectfully disagree that our work is primarily engineering-focused. The key contribution lies in the novel text-only RL training for VLMs to tackle a fundamental challenge: enhancing reasoning of VLMs without relying on expensive image-text paired training data.

We propose a text-driven reinforcement learning framework that uses language as the primary training medium to develop transferable reasoning capabilities, which are then effectively applied to multimodal decision-making tasks during inference. This framework enables us to decouple reasoning from perception, allowing for scalable and cost-efficient training while preserving strong performance in vision-grounded scenarios.

By demonstrating that such reasoning skills can be learned purely from text and transferred to complex visual tasks, our work offers a new conceptual perspective on VLM training that goes beyond the use of existing RL methods, and addresses a practical gap in current VLM development.

Thank you for your continued engagement in the rebuttal process. We are encouraged that you like our idea of text-only training to solve the visual-language gap.

To further address your concern raised in the original comments that “the performance of GRPO may rely heavily on the backbone VLM (Qwen2.5-VL as a strong open-source model),” we conduct additional experiments using Qwen2-VL [1], a base model known to have weaker language reasoning abilities [2]. This helps evaluate the robustness and generality of our approach across models with varying capacities.

Specifically, we apply our text-only RL training to Qwen2-VL-7B-Instruct. Due to time and resource constraints, we focus on Stage-2 training only, without the math cold-start phase. The results are as follows:

• Note: Due to time constraints, we evaluate on EgoNormia-verified, a 200-video high-quality subset provided by the original paper and leaderboard.

As shown, our method consistently improves performance across all benchmarks, even when applied to a base model with more limited reasoning ability. This further supports the effectiveness and generality of our approach.

[1] Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution

[2] Qwen2.5 Technical Report

Thank you for the thoughtful feedback. We are encouraged to know that you consider our work sound, interesting, and valuable for scenarios where paired image-text data is scarce. Below we address the weaknesses and questions you raised.

W1 & Q1. Experiment with Qwen2.5-VL

We appreciate your suggestion and would like to clarify our choice of Qwen2.5-VL (3B and 7B).

First, due to computational constraints, we focus on models at the 3B and 7B scales. Qwen2.5-VL is selected because it is a strong, open-source VLM with competitive performance, offering a reproducible and reliable foundation for our study. Recent work [1,2] suggests that RL mainly elicits and refines knowledge embedded in the base model, rather than learning entirely new capabilities. Therefore, tackling complex decision-making depends critically on the strength of the base model, and Qwen2.5-VL models provide the best open-source option at this scale when our work is conducted.

Second, we follow established practice in reasoning VLM research, such as OpenVLThinker [3], R1-OneVision [4], and MM-Eureka [5], all of which also build upon Qwen2.5-VL. This ensures comparability and consistency with existing methods.

We agree that extending Praxis-VLM to larger scales or other architectures is an important future direction. However, our primary goal in this work is to establish the principle and effectiveness of text-driven reinforcement learning for improving reasoning and decision-making capabilities. We demonstrate that the reasoning skills learned through language-only training can be successfully transferred to multimodal inference. We will clarify this motivation and the rationale behind our current model choice more explicitly in the revision.

[1] Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?

[2] S1: Simple test-time scaling

[3] OpenVLThinker: An Early Exploration to Complex Vision-Language Reasoning via Iterative Self-Improvement

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

[5] MM-Eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning

---

W2 & Q2. Regarding descriptions of images

Thank you for raising this question. We would like to clarify the following points and address any potential misunderstanding:

W3. How are detailed descriptions imposed using GPT-4o?

Praxis-VLM does not use generated textual descriptions of images, either during training or evaluation.

Our approach is motivated by the scarcity of large-scale image-text paired datasets of vision-grounded decision-making for VLM training. To address this, we explicitly disentangle reasoning from visual perception and train Praxis-VLM entirely on a synthetic, text-only dataset, where situations, questions, and potential decisions, are fully described in natural language, as shown in Figure 3. These synthetic textual training data are generated by GPT-4o without any reference to benchmark images, ensuring no concept leakage.

During training, only the language model component of the VLM is updated, with the visual encoder disabled. We apply reinforcement learning with adaptive rewards on the synthetic text-only data to develop the model’s multi-step reasoning and decision-making abilities.

At inference time, the model directly processes the actual visual inputs (images or video frames) from benchmark datasets, without relying on any textual description from GPT-4o. The central idea is that the decision-making skills learned from text-only training transfer effectively to multimodal inference, avoiding the need for paired image-text training data or intermediate descriptions.

Q2. How can one be sure that a textual description of an image generated by GPT-4o is sufficient and accurate?

We would like to clarify that GPT-4o generated image descriptions are only used in our preliminary analysis (Section 2) to explore whether VLMs can perform decision-making when visual contexts are replaced by text descriptions. In these experiments, images are replaced with GPT-4o-generated or human annotated descriptions, and only the language model component of the VLM is used for prediction. These experiments indicates that decision-making and reasoning can be disentangled from visual perception, motivating our text-only RL training approach for improving VLM reasoning purely within the language space.

Since Praxis-VLM always uses actual visual inputs (i.e., real images) during benchmark evaluations, the accuracy or completeness of GPT-4o-generated descriptions is not relevant to the core method. Our framework remains fully grounded in the visual modality during inference, ensuring valid and reliable evaluation in vision-grounded decision-making contexts.

We will revise our paper to better clarify this.

---

Q3: "Major." vs. "Pass@1" Scores for Qwen2.5-VL-7B

This is an excellent observation! The "Major." score reflects the accuracy of the most frequent answer across 8 sampled outputs, while "Pass@1" measures whether at least one of the 8 outputs is correct.

For the base Qwen2.5-VL-7B model, these scores are close because the model predicts answers in a highly deterministic manner, without explicit reasoning. Its output distribution is sharp and concentrated, and the model tends to repeatedly generate the same answer. As a result, the most frequent answer ("Major.") almost always matches the "Pass@1" result.

In contrast, reasoning VLMs (Reason SFT & Praxis-VLM) make predictions with explicit reasoning, enabling them to produce a broader variety of reasoning trajectories, exploring diverse solution paths, before reaching the final answer. This diversity increases the likelihood that at least one of the 8 samples is correct, even if the most frequent answer might be wrong, thereby creating a larger gap between "Major." and "Pass@1" scores. We will revise our paper to make this clearer.

---

Regarding Figure 1 Location

Thanks for your suggestion. We will revise our paper and relocate the Figure 1 accordingly.

Thank you for your continued engagement in the rebuttal process. We appreciate your recognition that our underlying idea of using solely text data to enhance visual reasoning is interesting.

To further address your remaining concern regarding experimental justification, we have conducted additional experiments using Qwen2-VL [1], a base model known to have weaker language reasoning ability compared to Qwen2.5-VL [2]. This allows us to better examine the robustness of our method across models with varying capacities.

Specifically, we apply our text-only RL training to Qwen2-VL-7B-Instruct. Due to time and resource constraints, we focus on Stage-2 training only, without the math cold-start phase. The results are as follows:

• Note: Due to time constraints, we evaluate on EgoNormia-verified, a 200-video high-quality subset provided by the original paper and leaderboard.

As shown, our method consistently improves performance across all benchmarks, even when applied to a base model with more limited reasoning ability. This further supports the effectiveness and generality of our approach.

[1] Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution

[2] Qwen2.5 Technical Report

Thank you for the insightful feedback. We are encouraged that you find our work to be "nice and data-efficient," and appreciate the motivation and performance of the proposed method. Below, we address your questions:

W1 & Q1. Data Generation

Scalability of Data Pipeline

Our dataset pipeline is designed to be scalable and effective to create a challenging and diverse dataset that fosters multi-step reasoning: (1) For diversity, we adopt a batch generation strategy, prompting GPT-4o to produce 10 samples at a time. Each batch undergoes deduplication based on word-overlap similarity to remove highly similar situations; (2) For difficulty, we embed explicit complexity instructions in our prompts (Line 649), guiding GPT-4o to generate questions that require non-trivial deliberation.

Beyond these automated steps, we intentionally avoid manual filtering or curation, enabling a fast and scalable pipeline for dataset creation.

The manual seed effort is minimal: we craft only 10 seed questions (similar to Figure 8) to provide GPT-4o as in-context examples. These seeds serve only to establish task format and structure rather than domain-specific knowledge. Consequently, the manual intervention is minimal, enabling easy adaptation to new domains.

Diversity and Representativeness of Synthetic Data

To assess the diversity of the generated scenarios, we prompt GPT-4o to cluster the generated textual situations into topical categories. The results indicate a broad coverage across varied scenarios:

• Workplace Performance and Personal Issues,
• Resource Allocation
• Project Management
• Balancing Competing Interests
• Policy, Rules & Enforcement
• Ethical Dilemmas
• Interpersonal Conflict
• Emergency Handling
• Navigating Setbacks
• Event Planning & Logistics
• Balancing Inclusivity and Majority Preference

While the generated situations are more close to the domain of VIVA (human-centered situations), there can be a domain gap compared to other benchmarks (e.g., PCA-Bench’s embodied robotics or EgoNormia’s egocentric video contexts). However, Praxis-VLM demonstrates consistent improvements and generalization across all benchmarks, indicating that our text-driven RL training learns generalized reasoning skills rather than overfitting to domain-specific patterns. This is also supported by the analysis in Figure 5, showing that with our text-only RL training, Praxis-VLM successfully acquires fundamental and adaptable reasoning abilities for decision making.

In future work, we plan to explore advanced data selection approaches to further enhance model training efficiency.

---

W2 & Q2 & Q3. Regarding Visual Inputs Ablations

We appreciate the question and would like to clarify a possible misunderstanding. Large-scale multimodal (image-text paired) datasets for vision-grounded decision-making are scarce and, to our knowledge, no such open-source training data are currently available. To address this, our approach deliberately disentangles reasoning from visual perception, using text-only training to isolate language as the medium for learning transferable reasoning skills. At multimodal inference, where image-text benchmarks are available, the model applies these reasoning abilities directly to real visual inputs (as shown in Figure 3).

W2 & Q2 & Q3. Why not use image captions?

Because large-scale multimodal training data is not available for this task, we cannot explore replacing our synthetic textual scenarios with image captions. This limitation is precisely why we investigate whether reasoning skills learned from pure text training can transfer effectively to multimodal settings, bypassing the need for expensive multimodal training data.

Q3. Visual vs. Text Reasoning Ablation

We cannot test training with image captions due to the lack of multimodal training datasets to obtain such captions. However, our method is inherently compatible: If we can generate captions, they could serve as additional textual situations to our synthetic data. Our current results already demonstrate that text-only training transfers effectively, as Praxis-VLM shows consistent improvements across multimodal benchmarks without any image-text multimodal training.

---

W3 & Q4 - Q6. RL Training and Reward

Q4. Increasing reasoning length does not reduce accuracy

We would like to clarify that our analysis (Sec 5.2) shows that Praxis-VLM tends to generate longer reasoning for more challenging samples. The observed accuracy drop is therefore linked to sample difficulty rather than the reasoning length itself. We will clarify this in the revision.

W3 & Q4. Ablation of $R_{len}$

Besides accuracy reward, we follow DeepSeek-R1 to include the format-related rewards ($R_{tag}$, $R_{format}$) to ensure outputs follow the explicit reasoning and prediction structure. We incorporate the length reward ($R_{len}$) to encourage more comprehensive analysis and reasoning, which is particularly beneficial for complex decision-making tasks.

To quantify its impact, we conduct an ablation study on Praxis-VLM-7B by removing $R_{len}$. The results show that $R_{len}$ consistently improves performance across benchmarks:

These results confirm that longer and more deliberate reasoning chains helps the model perform deeper analysis, leading to accuracy gains.

Q5. Contributions of text-to-text reasoning vs. GRPO training

We clarify that the text-to-text reasoning and the GRPO training are fundamentally interdependent. Praxis-VLM is trained with the GRPO algorithm on synthetic text-to-text data to develop reasoning abilities, which are then transferred to multimodal inference. Thus, we cannot isolate the effects of text-to-text reasoning and GRPO training independently.

W3 & Q6. Novelty of Method

While our approach shares the general structure of multi-stage RL, the novelty of our adaptive reward lies in its alignment with stage-specific reasoning skills and its application to text-driven training for multimodal decision-making.

First, unlike prior GRPO setups that rely on paired image-text data, we apply this adaptive reward strategy in a text-only RL setup, with the goal of transferring learned reasoning skills to vision-grounded inference.

Second, our reward function is explicitly tailored to target different cognitive competencies progressively: Stage 1 focuses on format adherence and general logical reasoning; Stage 2 shifts the reward signal toward situational reasoning and decision making.

This enables data-efficient learning and effectively equips the model with decision-making capabilities, without requiring costly multimodal supervision.

---

W4. Limited Performance Gains

We acknowledge your concern and highlight the practical significance of our model from two perspectives.

First, all benchmarks are recently proposed and designed to challenge current VLMs. While the gains on in-domain benchmarks such as VIVA are around 3%, it achieves +13% over the baseline on PCA-Bench (7B) and +8% on EgoNormia (7B), demonstrating its ability to handle diverse, out-of-domain scenarios. Moreover, Praxis-VLM is trained in a data-efficient manner, without requirement of image-paired data, reinforcing its practical utility.

Second, we further compare Praxis-VLM with OpenVLThinker [1], a recent reason-VLM based on Qwen2.5-VL-7B for general visual reasoning trained with multimodal data.

These results highlight that with text-only training, Praxis-VLM demonstrates superior performance across all tasks.

As Praxis-VLM shares the same parameter count as Qwen2.5-VL and is trained solely with text-only data, it introduces minimal added complexity while delivering practically meaningful gains.

---

W5. Cold-start Initialization

The math cold-start stage aims to equip the VLM with abstract, generalizable logical reasoning before Stage 2 task-specific training. This initialization proves crucial for cross-domain generalization, particularly on the EgoNormia dataset, where Praxis-VLM achieves notable gains.

Meanwhile, the smaller improvements on VIVA are mainly due to domain alignment: our synthetic text corpus in Stage 2 is similar to VIVA’s human-centered tasks (as discussed in W1 & Q1), so the benefits of math pretraining are less pronounced. We will revise our paper (Line 247) to better clarify this.

---

W6 & Q7. SFT Performance

The degradation of SFT-based models mainly arises from behavioral cloning and overfitting (Line 235). As discussed in W1 & Q1, there exists domain gaps between synthetic data and certain benchmarks like PCA-Bench and EgoNormia datasets. SFT training tends to enforce pattern memorization from the training distribution, which may hurt performance when the model is tested on benchmarks from divergent domains.

This observation aligns with prior work [1,2] that reason with SFT primarily establishes reasoning structures without encouraging exploration or adaptability. In contrast, RL training allows models to explore beyond static dataset examples, facilitating more fundamental and transferable decision-making skills. This difference explains why our GRPO-trained Praxis-VLM consistently generalizes better compared to its SFT counterparts.

[1] OpenVLThinker: An Early Exploration to Complex Vision-Language Reasoning via Iterative Self-Improvement

[2] SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training

Dear Reviewer J55j,

Thank you again for your valuable feedback and suggestions on our work. We have carefully addressed your comments regarding synthetic data generation, visual input analysis, ablations on RL rewards, model performance, and further clarifications on cold-start initialization and SFT result analysis.

As the discussion period will conclude in two days, we kindly invite you to review our response. We hope it sufficiently addresses your concerns, and we remain open to any further questions you may have.

Best regards,

The Authors

Thank you for your valuable comments. We appreciate your recognition of our work's innovative method, effectiveness and robustness of the model, and thoughtful analysis. Here we address your concerns and questions:

W1: Regarding Multiple-Choice QA Format

This is a good question! We choose multiple-choice question (MCQ) benchmarks for two primary reasons:

1. Objective and Reproducible Evaluation: The MCQ format allows for a straightforward and consistent evaluation of decision-making with accuracy, which has been widely adopted by prior work on reasoning-based VLMs [1,2]. In our case, it serves as a stable testbed for evaluating final decisions while avoiding the potential bias or ambiguity in scoring open-ended outputs. Evaluating open-ended actions, especially in complex, real-world scenarios, remains a challenging task. Therefore, we focus on MCQ based benchmarks for model evaluation.
2. Challenging and Diverse Scenarios: The benchmarks we used are recently proposed and specifically designed to evaluate decision-making in a wide range of complex, vision-grounded situations. They span human-centered social contexts, embodied agent tasks (e.g., robotics and autonomous driving), and egocentric video understanding, offering a comprehensive evaluation suite for our methods.

We acknowledge that extending these learned skills to interactive or open-ended environments remains an important and exciting direction for future work. Our current study, however, establishes a critical and reproducible step toward developing models capable of such real-world applications. We will revise our paper to clarify the benchmark choice.

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

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

---

Q1 & W2: Regarding Analysis on Visual Feature

Q1. Model Performance when image descriptions are imperfect

We first clarify that Praxis-VLM does not use image descriptions during inference. The generated image descriptions are only used in our preliminary analysis (Section 2) to explore whether VLMs can perform decision-making when visual contexts are replaced by text.

To assess the effect of descriptions on VLMs, we conduct an additional experiment where we replace the high-quality image descriptions used in our preliminary analysis (Figure 2, generated by GPT-4o or dataset annotations) with descriptions generated by a weaker model, LLaVA-1.5-7B. We then test the original Qwen2.5-VL-7B model's accuracy.

The results show a significant performance drop when using lower-quality, noisy descriptions, confirming that robust situational understanding is crucial for reliable decision-making to reach a good score for the benchmarks in our experiments.

These findings indicate that the vision encoder's quality remains a critical factor in the model’s inference-time reliability. However, enhancing visual perception typically requires paired image-text training data, which is unavailable in our target scenarios and beyond the scope of our work. Our approach addresses this by disentangling reasoning from perception, allowing reasoning to be trained efficiently on scalable text-only data.

W2. Does Praxis-VLM Actually Use Visual Features or Relies on Prior Learned Textual Knowledge?

This question is related to our discussion above. First, the significant performance drop when using noisy image descriptions (LLaVA-generated text) indicates that accurate visual perception is essential for VLMs to achieve strong performance on these benchmarks: when situational descriptions degrade, overall performance declines substantially. Praxis-VLM’s strong results on all three benchmarks suggest that the model is not applying patterns learned from text-only training, but instead actively leveraging visual information during inference.

Second, as detailed in our reasoning analysis (Section 5.4), a key component of the model’s reasoning is Situational Analysis, where the model explicitly interprets and grounds its reasoning in visual content. This is further illustrated in the output examples (Figures 10–12), where the model consistently begins its reasoning by analyzing the visual situation. These findings collectively suggest that Praxis-VLM grounds its reasoning in visual evidence, rather than relying solely on text-trained patterns. We will highlight this conclusion in our revision.

We will revise our paper to better clarify this.

---

Q2. Regarding Synthetic Data Quality and Coverage

Our synthetic dataset pipeline is designed to be scalable and effective to create a challenging and diverse dataset that fosters multi-step reasoning: (1) For diversity, we adopt a batch generation strategy, prompting GPT-4o to produce 10 samples at a time. Each batch undergoes deduplication based on word-overlap similarity to remove highly similar situations; (2) For difficulty, we embed explicit complexity instructions in our prompts (Line 649), guiding GPT-4o to generate questions that require non-trivial deliberation.

Beyond this, we avoid manual filtering or heavy curation, enabling fast and domain-agnostic data generation.

To assess the diversity of the generated scenarios, we prompt GPT-4o to cluster the generated textual situations into topical categories. The results indicate a broad coverage across varied scenarios:

• Workplace Performance and Personal Issues,
• Resource Allocation
• Project Management
• Balancing Competing Interests
• Policy, Rules & Enforcement
• Ethical Dilemmas
• Interpersonal Conflict
• Emergency Handling
• Navigating Setbacks
• Event Planning & Logistics
• Balancing Inclusivity and Majority Preference

Notably, while the generated data cover diverse situations, there can still be a domain gap when compared to certain evaluation benchmarks (e.g., PCA-Bench’s embodied robotics or EgoNormia’s egocentric video contexts). However, Praxis-VLM’s consistent improvements and strong generalization across these benchmarks indicate that the text-based training equips the model with generalized and transferable reasoning skills. Figure 5 also shows that Praxis-VLM learns general reasoning dimensions (situational analysis, consequence evaluation, safety, and norm adherence), enabling effective transfer to multimodal scenarios.

However, if the text data lacks variety or introduces bias, we believe the model performance will drop. Unlocking fundamental decision-making skills requires a wide variety of situations and perspectives. Prior work also confirms that dataset diversity is a key factor influencing model performance and alignment [1, 2]. Moving forward, we plan to explore advanced data selection methods to further enhance effective model training.

---

Q3. Role of Math Pretraining

This is a great and insightful point! Our motivation for incorporating the math cold-start stage is to provide the model with a foundation in abstract, generalizable logical reasoning before introducing task-specific decision-making in Stage 2. The results show that this cold-start initialization measurably enhances the model’s generalization capabilities, particularly on the out-of-domain EgoNormia benchmark (as discussed in line 247). This demonstrates that the math training strengthens the model’s underlying reasoning and improves its adaptability to novel, complex decision-making scenarios.

This finding is also consistent with concurrent work [1], which shows that including mixed-domain training data can improve general reasoning performance. We will revise our paper to better clarify our motivation.

Measurable Improvements in reasoning

This is a good question! Quantifying reasoning depth and structure for comparison remains a challenge due to the lack of standardized metrics. To approximate this, we utilize GPT-4o as a Rhetorical Structure Theory (RST) discourse parser and convert each generated reasoning into a discourse tree. The average RST tree depths are comparable across models (3.93 vs. 3.90 for models with and without math pretraining), indicating similar surface structure.

Therefore, we further conduct a manual inspection on examples where only the pretrained model produces correct predictions. We find that Praxis-VLM with math pretraining integrates more relevant symbolic visual cues into its reasoning, particularly in complex contexts. For instance, in PCA-Bench's autonomous driving scenarios, Praxis-VLM with math pretraining is able to identify and integrate multiple road symbols (e.g., traffic signs, lane markers) into its reasoning, leading to more contextually appropriate action choices, whereas the variant without math pretraining tend to rely on a narrower subset of cues and make less contextually appropriate action choices.

We hypothesize that the symbolic reasoning rigor inherent in math and geometry tasks foster a more systematic reasoning framework. This structured reasoning appears to transfer effectively to multimodal decision-making by promoting careful observation, multi-step analysis, and integration of fine-grained visual signals.

We will revise the paper to clarify this motivation and highlight these empirical insights more explicitly.

[1] LIMA: Less Is More for Alignment

[2] Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective

Dear Reviewer jthP,

Thank you very much for your thoughtful and constructive feedback, which has been invaluable in helping us improve our work. We have carefully addressed your comments regarding the question format, analysis of visual feature usage, synthetic data generation, and clarifications on cold-start initialization.

As the discussion period concludes in two days, we kindly invite you to review our response. We hope our response adequately addresses your concerns, and we remain happy to clarify any remaining questions you may have.

Best regards,

The Authors