See&Trek: Training-Free Spatial Prompting for Multimodal Large Language Model

We thank the reviewer for their thoughtful comments. We are glad that the reviewer appreciated the practical relevance and timeliness of addressing spatial reasoning in vision-only MLLMs, and the intuitive design supported by thorough evaluations. We now address your comments point by point below.

1. Concern about Related Work Discussions. These works have made important contributions to training-free visual prompting paradigms. Here, we compensate for these discussions with these works as follows: D2R [1] is a training-free prompting framework, which integrates textual CoT with corresponding visual drafts into MLLMs. ViCrop [2] leverages a MLLM’s own internal attention and gradients to automatically crop images, significantly improving its perceptual accuracy on small visual details that it would otherwise fail to recognize. ControlMLLM [3] enables MLLMs to follow specific visual prompts by optimizing a learnable latent variable at inference time to control the model's attention. In addition to these works, we also discuss some prompting works that are beneficial to enhancing spatial understanding ability in the appendix Section B.6., such as CC [4] or SoM [5]. Different from these works, See&Trek aims to enhance the spatial understanding of MLLMs under vision-only constraints in a training & GPU-free manner, which focuses on two core principles: increasing visual diversity and motion reconstruction.

2. Concern about Each Component Performance. We appreciate the reviewer’s careful inspection of our ablation results (Table 3/4/5 in main papers).

• a. Compared to Temporal-TopK, Balanced-TopK focuses more on semantic richness, which selects the frame with the most classes within each temporal segment in a global view. Thus, Balanced-TopK can cover object classes in the video as much as possible. The diversity of the collected scenes helps MLLM understand the spatial relation between objects, namely, the spatial structure in the video. The experiment in Table 6 provides evidence of the hypothesis: Balanced-TopK achieves consistent gains on several key spatial reasoning tasks compared to Temporal-TopK like 1) Relative Distance: +4.9% 2) Room Size: +1.2% 3) Absolute Distance: +1.4% 4) Relative Direction and Approximate Order: +0.5% each. However, Balanced-TopK will ignore the frame with fewer objects whose classes are already collected. If these objects are not in the previously collected frames, MLLM will never be aware of them. Thus, we observe a performance decrease in the metrics related to objects themselves, e.g., Object Count.

• b. Each module is designed to target specific spatial reasoning capabilities, rather than boosting all tasks uniformly. we would like to emphasize that the design of our method explicitly targets different subtypes of spatial reasoning, and the ablation tables (Table 3 and 4) demonstrate this alignment:

  • MSRS focuses on increasing semantic diversity, which is reflected in its gains in Approximate Order (+3.6%) and Relative Direction (+6.6%)
  • Motion Reconstruction enhances the model’s understanding of trajectory and transitions, leading to significant improvements in Route Plan (+2.5%) and Approaching Order (+14.7%).
  • Spatiotemporal Encoding and Point Prompts aim to connect visual diversity component and motion construction, offering complementary spatiotemporal cues, showing targeted gains in Abs. Distance (+1.9 ~ 2.2%), Relative Distance (+1.7~2.0%), and Route Plan (+4.1%). It demonstrates that these design can aid in temporal alignment and enhance geometric grounding.
  • The ablation study of the internal component design and the corresponding experiments can also be found in Appendix like Table 7-10. We hope that this additional experimental information can resolve your doubts.

3. Concern about More Baseline Models. We thank the reviewer for pointing out the inconsistency between the number of models reported in Table 1 (VSI-BENCH) and Table 2 (STI-BENCH). We acknowledge that in our original submission, only two open-source models were reported on STI-BENCH due to early-stage computational resource constraints. This was not intended as cherry-picking but rather a reflection of practical limitations during initial benchmarking. To address this concern, we now conduct a comprehensive evaluation of some previously tested MLLMs on STI-BENCH and incorporated the complete results into the revised version. As shown in the updated Table below, SEE&TREK consistently improves performance across most models, with gains ranging from +1.0% to +1.4% on average.

4. Concern about the Closed-source Models Performance. We appreciate the reviewer’s feedback regarding the evaluation scope across MLLMs, including proprietary models. Here, we report results for proprietary models (GPT-4o, Gemini-2.0-Flash) on VSI-BENCH.

As shown above, these commercial engines benefit from our SEE&TREK framework across various spatial tasks, reinforcing its generality and practical effectiveness.

5. Concern about Improvements in Stronger Models. We acknowledge that as model capacity increases, the relative gains from external input modifications may diminish. This is a natural outcome of the ceiling effect, where high-capacity models already encode stronger spatial priors and benefit from richer pretraining corpora. Nonetheless, we argue that even marginal improvements (e.g., +1.2% on Qwen2.5-VL-32B) are meaningful, especially in the zero-shot, training-free setting of our method, which requires no fine-tuning, no gradient updates, and no GPU.

6. Concern about the Framework Generalization. We thank the reviewer for raising this important point regarding generalization.

• Benchmark choice reflects diversity, not task specialization. We fully agree that generalization is a crucial criterion for evaluating practical methods. While our experiments are centered on VSI-BENCH and STI-BENCH, these benchmarks are deliberately chosen not for narrow task matching but for their breadth and diversity (We also have stated these benchmark details in appendix Section B). Specifically, these benchmark include real-world egocentric videos across indoor, desktop, and outdoor scenarios, requiring models to reason about object relations, sizes, and topology across diverse camera viewpoints and occlusion patterns. They provide a rigorous and multi-perspective evaluation of spatial understanding that goes beyond task-specific tuning.
• Our approach is not designed to exploit dataset-specific structures but instead targets two fundamental bottlenecks of MLLMs in video-based spatial reasoning: 1) Visual Homogeneity, addressed by the Maximum Semantic Richness Sampling module. 2) Unknown Camera Motion, addressed via: a classical VO-based Motion Reconstruction module for trajectory estimation and Spatiotemporal Encoding, which overlays motion cues and temporal indices onto keyframes to bridge semantics and motion. Each component within the framework is modular and can be easily adapted or substituted to meet different domain constraints and resource requirements. We have perform some experiments like utilizing some SOTA components as shown in R3 Response 3 to verify our effectiveness.
• We also provide the OpenEQA results for comprehensive evaluation here. The results demonstrate the superiority of our spatial prompting framework design.

7. Concern about the Ablation Study. Thanks for your insightful comments. We have performed a comprehensive ablation study as shown in Table 3-6 in the main paper. Moreover, we also make a deep investigation into each component design or hyper-parameter setting as shown in Table 7-10 in the appendix. We wish these experiments information can address your concerns regarding ablation.

8. Concern about the Scope. We are agree and thank you for the thoughtful comment. Our work focuses on spatial understanding from purely visual video inputs, without relying on depth or other modalities. Compared to image-based tasks, video-based reasoning is more challenging due to temporal dynamics and spatial consistency. By leveraging motion cues and scene continuity, our method enables richer spatial understanding and highlights the practical value of video-centric, training-free prompting.

Finally, thanks for your constructive feedback. Hope that we have carefully addressed all your concerns above, and we sincerely appreciate your thoughtful contribution you’ve made to improving our work. We will incorporate these revisions into the revised version.

[1] Bridging the Dynamic Perception Gap: Training-Free Draft Chain-of-Thought for Dynamic Multimodal Spatial Reasoning.

[2] MLLMs Know Where to Look: Training-free Perception of Small Visual Details with Multimodal LLMs.

[3] Controlmllm: Training-free visual prompt learning for multimodal large language models.

[4] COARSE CORRESPONDENCES Boost Spatial-Temporal Reasoning in Multimodal Language Model

[5] Set-of-Mark Prompting Unleashes Extraordinary Visual Grounding in GPT-4V

We thank the reviewer for their thoughtful comments. We are glad that the reviewer appreciated the practical value of our method, particularly its training-free nature, low resource requirements, and effectiveness in enhancing spatial reasoning in MLLMs using only visual inputs. We now address your comments point by point below.

1. Concern about the Technique Combination. We appreciate the reviewer’s comment and would like to clarify that, the conceptual motivation and problem framing behind SEE&TREK are novel and underexplored in the current literature, even though our method is developed based on existing components such as object detection and visual odometry. As we emphasize in the Introduction, prior works on enhancing MLLMs’ spatial understanding often rely on annotated spatial semantics (e.g., number cues, axis or semantic masks), or require model retraining and architectural modifications (We also discuss these work in appendix Section B.7. ). However, how to stimulate spatial reasoning in MLLMs under a purely visual and training-free setting remains an open challenge, which is beneficial for most areas like robotics. In our work, we take a step back and reflect critically on this problem setting, Iidentifying two major and fundamental bottlenecks: Visual Homogeneity and Unknown Camera Motion. Building upon these observations, SEE&TREK is designed from the ground up to explicitly address these two limitations through visual diversity and motion reconstruction. While the techniques we use (e.g., YOLO, VO) are established, their roles in our pipeline are not trivial. These components are not mere add-ons but are core to realizing our conceptual framework. Their interplay—semantic-aware selection plus trajectory-aware encoding—allows SEE&TREK to effectively emulate spatial continuity and dynamics from static visual inputs, offering a principled way to address current MLLM limitations. In this sense, SEE&TREK has a novel perspective on designing spatial prompts that help MLLMs better perceive and reason in 3D space, under a resource-efficient and training-free regime. Importantly, the current implementation presented in this paper represents a minimal working instantiation of this conceptual framework—i.e., GPU-free or maximum accessibility and deployment ease. Moreover, the framework itself is flexible and modular: depending on the application scenario, the individual components can be upgraded. For example, more advanced feed-forward 3D networks [1,2] or VO like MAC-VO [5] could be employed to capture holistic scene geometry. Here, we provide some results on VSI-Bench by utilizing diferent SOTA perception models in Following Response 3.

2. Concern about Experiment Performance. Regarding the negative performance changes observed on specific tasks or models: We thank the reviewer for raising this important point. We address this concern through both systematic trend analysis and qualitative case studies.

• (1) Systematic Trend Analysis (from Table 6): As noted, some tasks (e.g., Object Count, Object Size, Relative Direction) show minor drops in accuracy for certain models. We attribute this to limitations in our current module choices, which were intentionally selected for their GPU-free and lightweight nature, but may introduce trade-offs in robustness and granularity:
  • a. VO module drift: We use the manual Visual Odometry module, which offers fast and training-free pose estimation. However, it is known to suffer from accumulated drift, especially in high-speed or large-amplitude camera motions, which are commonly found in datasets like VSI-BENCH and STI-BENCH. In such dynamic scenes, the VO module may fail to maintain globally consistent pose trajectories, leading to incorrect relative position estimates. This directly impacts tasks such as Relative Direction, Route Plan, and Room Size, where spatial coherence is critical.
  • b. Semantic sampling bias: Our default perception module, YOLOv8-tiny, is optimized for speed but has limited object coverage and detection recall. In complex scenes with multiple small, overlapping, or low-salience objects, this model may miss relevant entities, degrading performance in Object Size tasks.
  • c. Model capacity: Lightweight MLLMs (e.g., Qwen2.5-VL-3B) may struggle to understand motion-aware visual prompts from multiple frames due to its limited model scale, especially when spatial signals are sparse or noisy. This affects multi-step reasoning tasks such as Room Size or Relative Distance, where consistency over multiple frames is essential.
  • d. Missing frames: The proposed MSRS tends to collect the scenes with the most object class diversity. It sometimes ignores the frame with fewer objects whose classes are already collected. If these objects are not in the previously collected frames, MLLM will never be aware of them. Thus, we observe a performance decrease in the metrics related to objects themselves, e.g., Object Count.
• (2) Qualitative Case Studies. We have also conducted some failure case analysis to better understand these limitations. As illustrated in Appendix Figure 9, SEE&TREK occasionally fails even when all queried objects are visible, due to misalignment between egocentric motion cues and layout inference (e.g., in route plan or relative distance tasks). This reinforces our motivation to explore more powerful dense perception models in future work to boost semantic richness under diverse spatial contexts.

3. Concern about Generality across Different Detection or Motion Estimation Pipelines. Thanks for the practical comments. This paper mainly proposes a general and training-free framework that explicitly incorporates semantic-rich frame sampling and motion reconstruction to enhance spatial reasoning. To validate this idea in a lightweight and accessible manner, we implement a minimal working solution using YOLOv8-tiny and manual VO module, both of which are GPU-free and efficient. This implementation demonstrates that even under constrained settings, our method can deliver consistent performance gains. Importantly, our framework is highly extensible. The perception and motion modules can be readily replaced with stronger alternatives according to different scenarios. Here, we conduct experiments on the VSI-Bench involving the replacement of various alternatives, such as substituting the detection component with VLM-based Models (Grounding Dino and Yolo-World) and the VO component like MAC-VO. Note that we utilize the InternVL3-8B+See&Trek as our baseline.

Finally, thank you for your thoughtful comments and constructive suggestions; we appreciate your recognition of our method’s strengths as well as your valuable insights on framework robustness and generality. We hope our response can solve your doubts.

[1] Vggt: Visual geometry grounded transformer

[2] Grounding image matching in 3d with mast3r

[3] Grounding dino: Marrying dino with grounded pre-training for open-set object detection

[4] Yolo-world: Real-time open-vocabulary object detection

[5] MAC-VO: Metrics-aware Covariance for Learning-based Stereo Visual Odometry

We sincerely thank Reviewer uJnW for the constructive feedback and for acknowledging our rebuttal. We are glad that our responses have addressed your concerns, and we will incorporate the discussed points—such as the VO failure analysis and generalization limitations—into the revised version to improve the clarity and completeness of the paper. Our team deeply appreciate your contribution throughout the review process.

We thank the reviewer for their thoughtful comments. We are glad that the reviewer found our method conceptually novel, well-written, clearly structured and technically sound, and also recognized our performance. We would like to address all of your concerns and questions below with point responses:

Concern about The Sampling Interval in Table 5. Thanks for your significant consideration. The absence of a clear trend across different sampling intervals mainly stems from the following reason:

1. our focus is on extracting semantically rich content, rather than dense temporal coverage. The Maximum Semantic Richness Sampling (MSRS) strategy ensures that keyframes are selected based on visual informativeness rather than fixed temporal intervals. Thus, even when the sampling interval $N$ varies, the retained frames tend to be semantically similar. In fact, using a very small interval (e.g., $N=1$) may introduce redundant or noisy frames, which can damage the model’s spatial understanding and reduce answer consistency. This is partially reflected in the slight performance dip at $N=2$ compared to $N=3$ or $N=4$.

2. In addition, the sampling interval $N$ has a substantial impact on processing time. As shown in Table 5, increasing $N$ significantly reduces the computational cost (e.g., from 410s at $N=1$ to 82s at $N=4$), while maintaining comparable or even better average accuracy. This indicates a favorable trade-off between efficiency and effectiveness, where moderate sampling (e.g., $N=3$ or $N=4$) strikes a good balance for most use cases.

3. More investigation and exerpeiment results about sample internval and detectors setting can be found in appendix Table 9.

Thank you for your kind words and positive assessment of our work. We also appreciate your insightful question regarding the sampling interval ablation in Table 5. We hope our response can solve your doubts.

We sincerely thank Reviewer nToN for the thoughtful feedback and positive evaluation. Your insightful question contributed to improving the completeness of our work, and we will incorporate the corresponding discussion into the final revised version. Your recognition is the greatest encouragement to our team.

We thank the reviewer for their thoughtful comments. We are glad that the reviewer found our method to have widespread applicability and the novelty of incorporating camera motion as an explicit cue. We would like to address all of your concerns and questions below with point responses:

1. Concern about Poor Writing and Notational Overload. Thanks for your thoughtful feedback. We are sorry for the confusing notation raised in Section 3. Here, we modify our notation for better readability. The $\tau_t$ denotes the selected keyframe index. Hence, the $f_{\tau_0}$ can represent the frame sampled in the initial phase. Then, the $\tau_g$ from Eq. (1) is re-written as $\tau_0$. Then, $\tau_i$ is rewritten as $\{f_{\tau_i}\}$ in Line 125, which represents the selected frames. Moreover, we also make the abstract algorithm flow as shown in Algorithm 1 in the main manuscript, and the detailed algorithm flow is as shown in the appendix Algorithm 2. We hope these modifications and other information can be helpful for our method formulation reading. Regarding Figure 2, we also are sorry about the visual clutter. We will make it better in the revised version.

2. Concern about Missing Related Work. Thanks for pointing out this important point. Regarding the Coarse Correspondence [1] (CC) or more similar prompting methods like SoM [2], We have discussed their difference and performed a comparison study in the Appendix Section B.7.

As shown in Table, See&Trek consistently outperforms existing visual prompting methods such as CC or SoM across all spatial reasoning tasks. While CC and SoM primarily rely on static masks or coarse positional labels to identify object regions, they offer limited support for modeling inter-object relationships or capturing scene-level spatial structures. In contrast, See&Trek leverages two key principles—visual diversity and motion reconstruction—to provide richer and more structured spatial cues.

3. Concern about Usage of Camera Intrinsics. We appreciate the reviewer’s concern. To clarify, we do not utilize the intrinsic parameters provided by the original datasets (e.g., VSI-BENCH or STI-BENCH). Instead, we just adopt a fixed intrinsic matrix derived from the KITTI dataset (K = [[718.8560, 0, 607.1928], [0, 718.8560, 185.2157], [0, 0, 1]]) across all experiments. This design choice was made deliberately to demonstrate the plug-and-play nature and extensibility of our framework, even in scenarios where intrinsic parameters are unavailable—a common case in real-world video sources. While using accurate intrinsics may lead to improved VO estimates and absolute metric precision, our method is built around relative motion encoding, not absolute scale recovery. To account for the inaccuracies introduced by this assumption, we explicitly emphasize in the prompt that the spatial information reflects relative distances and orientations, rather than real-world metric measurements. Lastly, our method is designed as a training-free spatial prompting module, intended to enhance the MLLM’s spatial understanding through motion-aware input construction. The use of intrinsics is strictly limited to internal VO estimation and does not influence or depend on whether baseline methods utilize such information. Therefore, the comparison remains fair, and the gains reflect the effectiveness of our spatial enhancing strategy.

4. Concern about Closed-source Models Performance. Thanks for your insightful comment. We perform the evaluation on the commercial engine GPT-4o and Gemini-2.5-Flash on VSI-bench set as follows: | Method | Gains | Avg. | Obj. Count | Abs. Dist. | Obj. Size | Room Size | Rel. Dist. | Rel. Dir. | Route Plan | Appr. Order | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | Gemini-2.0-Flash | - | 42.6 | 51.1 | 29.8 | 55.1 | 56.2 | 38.2 | 40.9 | 32.1 | 37.4 | | +SEE&TREK | +1.2% | 43.8 | 49.8 | 29.3 | 56.4 | 57.0 | 39.4 | 41.9 | 35.4 | 41.8 | | GPT-4o | - | 34.0 | 46.2 | 5.3 | 43.8 | 38.2 | 37.0 | 41.3 | 31.5 | 28.5 | | +SEE&TREK | +1.4% | 35.4 | 46.5 | 5.0 | 45.1 | 40.6 | 39.0 | 43.2 | 33.5 | 31.5 |

As shown in the table, our method consistently improves performance across different commercial engines.

5. Concern about Marginal Performance Gains. We thank the reviewer for highlighting it. It is important to emphasize that spatial reasoning remains an inherently challenging task for MLLMs, due to both the limited spatial supervision in pretraining corpora and the resource constraints in downstream evaluation (e.g., limited frame budgets in benchmarks). Even small performance gains (e.g., <2%) are meaningful in this context, as they indicate improved geometric and relational understanding under minimal input. Moreover, larger models tend to possess stronger built-in spatial priors, making additional improvements naturally more constrained. However, our method consistently yields positive gains across models of all scales, including powerful backbones like KIMI-VL-A3B, which demonstrates its general applicability.

6. Concern about the Benchmark Coverage Problem. We appreciate this important suggestion about coverage. we conduct the experiments on OpenEQA [3] with InternVL3-8B to show our method's superiority. Note that we utilize the EM-EQA setting [3] to evaluate the performance, and we still set 8 frames as input.

As shown in the Table, SEE&TREK brings consistent improvements across models, with a +1.8% gain on Qwen2.5-VL-7B and +2.0% on InternVL3-8B, demonstrating its effectiveness in enhancing spatial reasoning without fine-tuning.

While ScanQA and OpenEQA are valuable contributions to the spatial reasoning community, our choice of VSI-Bench and STI-Bench was deliberate and grounded in their superior diversity, task coverage, and spatial evaluation rigor, particularly when evaluating the spatial intelligence of MLLMs. First, VSI-Bench and STI-Bench are not limited to “low-level spatial perception tasks,” but instead represent high-level and fine-grained spatial reasoning benchmarks across diverse and realistic visual environments. These benchmarks are specifically designed to stress-test an MLLM's ability to understand fine-grained spatial structures, dynamic motion, and scene geometry, rather than relying on language priors or generic commonsense. ( We also have stated these dataset details in the Appendix. ) In contrast, while ScanQA or OpenEQA are earlier benchmarks for spatial QA, they tend to focus on relatively narrow task formats, predominantly involving object-level recognition, existence, and co-location in static scenes (ScanQA) or episodic memory from passive video (OpenEQA). Moreover, the evaluation metrics used in ScanQA or OpenEQA—BLEU, METEOR, ROUGE-L, CIDEr—are primarily adapted from traditional VQA and text generation. These metrics often struggle to meaningfully distinguish spatial reasoning quality, especially when answers differ in surface form but not semantics, or when tasks require metric estimation, relative positioning, or trajectory-level reasoning. By contrast, VSI-Bench and STI-Bench employ structured evaluation formats (e.g., spatial multiple choice, numeric prediction with relative error thresholds) and task-specific accuracy metrics designed to quantitatively and unambiguously measure spatial competence. This ensures that improvements in model performance—such as the +3.5% gain brought by See&Trek—are directly tied to enhanced spatial understanding, not merely improved textual fluency or memorization.

7. Concern about Computational Overhead. Thanks for highlighting this practical concern. Regarding the additional computational cost, we have the corresponding experiments on VSI-benchmark as shown in Line 276 ``Sample Efficiency Analysis'' in the main paper. To explore different YOLO's efficiencies, we also perform diverse experiments on Table 9 in the appendix. The conclusion is that the time consumed by SEE&TREK still largely depends on the length of the given videos. Particularly, the duration of videos in VSI-BENCH is at least longer than 1 minute.

Thank you again for your constructive feedback—we have carefully addressed all your concerns in our responses above, and we hope our clarifications help resolve your doubts. We are especially grateful for your thoughtful suggestions, which contribute meaningfully to advancing the spatial understanding of MLLMs.

[1] COARSE CORRESPONDENCES Boost Spatial-Temporal Reasoning in Multimodal Language Model

[2] Set-of-Mark Prompting Unleashes Extraordinary Visual Grounding in GPT-4V

[3] OpenEQA: Embodied Question Answering in the Era of Foundation Models

Dear Reviewer VBKe,

Thanks for your consideration and effort.

1. Regarding the randomized intrinsics, we will further organize our code and opensource it in github. And we also apprecate your investigation in COLMAP and consideration about raising our scores.

2. Regarding the performance difference between YOLOv8 and GPT-4o, it's difficult to quantify GPT-4o perception ability in traditional 2D detection benchmark. This prompts us to use detectors with the same architecture for comparison. However, GPT-4o's perception ability has been verified so significant in so many understanding task, which is the common sense.

We hope our response can solve your doubts.

Best Regards,

Team #5923

Dear Authors,

First, I want to clarify that I am not suggesting geometric priors cannot be used — for example, I have no issue with your use of YOLO. However, I consider camera intrinsics a different matter because they are often not easy to obtain in practice, unless (1) you have a simple and reproducible way to estimate them, or (2) their exact values do not meaningfully affect results.

You now claim that even with completely randomized intrinsics, RANSAC can still recover usable relative positions. I would appreciate it if you could share the exact randomized intrinsic matrices you used — ideally along with the small piece of code that generates them — so that I can better understand the degree of deviation from realistic parameters.

Regarding the GPT-4o point, I understand that See&Trek does not use GPT-4o, but in your pipeline YOLO is used only to detect the number of object categories in each frame and to sample keyframes accordingly. I do not believe replacing YOLO with GPT-4o for this specific task would yield a large performance boost — unless you can show that YOLO often selects incorrect keyframes with very few objects. For instance, in Table 3, MSRS improves by only 1.6%. Can you demonstrate that GPT-4o would significantly outperform YOLO in estimating relative object counts in a way that leads to a substantially better keyframe subset?

We sincerely appreciate the reviewer’s effort in engaging with our rebuttal despite their busy schedule. We would like to address all of your concerns and questions below with point responses:

• Concern about the Camera Intrinsics. Thanks for you consideration about the fairness.
  • The fixed KITTI intrinsics are used solely within the visual odometry (VO) stage to convert 2D correspondences into relative camera motion. Our prompting design explicitly frames all motion cues in terms of relative distances and orientations, without providing any metric calibration to the MLLM.
  • Due to time limitations during the rebuttal period, we can only provide a limited set of additional experiments. Nonetheless, we include results using normalized, random and identity intrinsics matrices in the VO stage to demonstrate that the performance gains persist even without any meaningful camera calibration. Here, we still adopt the InternVL3-8B + See&Trek as the baseline and evaluate on the VSI-Bench.

The results in the table demonstrate that replacing the original KITTI intrinsics matrix with normalized, random, or identity matrices yields comparable performance across all spatial reasoning sub-tasks on VSI-BENCH. The minimal differences in average scores and individual metrics indicate that our method’s effectiveness is not dependent on the accuracy or specific values of the camera intrinsics matrix $K$. This supports our claim that the spatial improvements stem primarily from the relative motion cues encoded by our VO and prompting strategy, rather than any geometric advantage conferred by precise intrinsic calibration. Consequently, our framework maintains robustness and fairness even when intrinsic parameters are unavailable or approximate.

• Concern about the Magnitude of Improvements. Thanks for given this work [1] for discussion. We believe the performance difference is largely due to the fundamental design trade-offs between the two approaches:
  • Our framework is deliberately designed to be training-free and lightweight, relying solely on prompting techniques. In contrast, [1] introduces the SpatialMind Prompt, which requires GPT-4o for frame-level perception filtering and scene modeling. GPT-4o, as a powerful commercial model, already exhibits very high baseline accuracy, which partly explains their larger absolute improvements. However, such an engine comes with significant network latency and large token consumption, limiting its applicability in resource-constrained scenarios. Our method, by contrast, employs skip-frame selection and a GPU-free perception module, which is easy to deploy, making it more suitable for a wider range of conditions.
  • Another key factor is that [1] also fine-tunes their VLMs on their curated dataset, as explicitly stated in their paper (“Fine-tuning VLMs on this dataset equips them…”). Fine-tuning directly injects spatial knowledge into the model parameters, which is inherently more powerful but also more resource-intensive and less generalizable. Our method does not involve any fine-tuning, preserving zero-shot applicability and enabling plug-and-play integration with different MLLMs without retraining.
  • Given these constraints and design goals, our performance gains—while numerically smaller—reflect improvements achieved under a much lighter computational and resource condition, which we believe is a valuable and complementary contribution to the field.

Once again, thank you very much for your constructive feedback. We have carefully considered and addressed all your latest concerns above, and we hope our clarifications provide useful insights to solve your doubts.

[1] Spatial Understanding from Videos: Structured Prompts Meet Simulation Data

Reviewer VBKe,

Please reply to the rebuttal and update the final decision.

AC