acceptodds
Under review as a conference paper at ICLR 2027

Spatial-OPSD: Self-Improving Spatial Reasoning via Label-Free Self-Distillation

Abstract

Vision-language models (VLMs) increasingly operate in embodied and spatially grounded settings, where accurate understanding of depth, viewpoint, and three-dimensional relations is essential. However, improving spatial reasoning typically relies on ground-truth answers, answer-derived rewards, or other forms of task-specific supervision. We introduce Spatial-OPSD, a label-free self-improvement framework that instead exploits spatial structure naturally available from perception and reconstruction tools. During training, a privileged teacher receives automatically obtainable spatial priors, such as depth, reconstructed 3D relations, and camera geometry, while the student observes only the original visual-language input. On trajectories sampled by the student itself, the teacher provides dense token-level supervision, allowing the student to internalize spatial knowledge without ground-truth answer labels or privileged information at inference time. To extend this supervision beyond a single round, we adopt a round-wise recursive training scheme: the teacher remains frozen within each round to provide a stable learning target, and the improved student initializes both teacher and student in the next round, where privileged spatial priors re-establish an informative teacher–student asymmetry. This enables repeated self-improvement while avoiding a rapidly moving teacher during optimization. Across four VLM families, a single round of Spatial-OPSD consistently improves the five-benchmark average, while three rounds further push a strong spatially specialized model to the open-source frontier, achieving the highest average among the open models and the best results on three of five spatial reasoning benchmarks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
vision-language models, spatial reasoning, self-distillation, multimodal reasoning, self-improvement

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CWfrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Spatial-OPSD for improving spatial reasoning in vision-language models through privileged-context on-policy self-distillation. A teacher initialized from the same checkpoint as the student receives question-specific geometric information, while the student sees only the original visual question. Training minimizes token-level reverse KL on student-generated responses without using task-answer labels in the objective. The teacher is frozen within each round and refreshed from the student between rounds. Experiments evaluate one-round training across four 4B model families and three-round training on SenseNova-SI-Qwen3-VL-8B, using five spatial benchmarks. Additional analyses examine teacher-refresh schedules, distillation objectives, and numerical noise in the spatial priors.

Critical strengths
  1. A useful application of privileged spatial information. Using native or estimated geometry to supervise a model without directly optimizing against task-answer labels is a practical direction. The resulting student requires no geometric tools or privileged inputs at inference.

  2. Improvements in some settings. Table 1 reports average gains of approximately 3.97 and 2.94 percentage points for Qwen3-VL-4B and InternVL3.5-4B, respectively. The matched three-round comparison in Table 3 also provides evidence that refreshing the teacher helps beyond simply extending fixed-teacher training on Qwen3-VL-4B/SPAR-Bench.

Critical weaknesses
  1. The contribution beyond existing OPSD is not sufficiently established. The method largely follows privileged-context OPSD, with question-specific spatial information replacing privileged solutions and the teacher refreshed between rounds. This is a reasonable adaptation, but the experiments do not clearly establish the value of the proposed training procedure over simpler ways of using the same geometric information. In particular, a relevant missing baseline is self-SFT: the same model receives the spatial priors, generates responses without access to GT answers, and those responses are used to fine-tune the student on the original inputs. The answer-supervised SFT and GRPO results in Table 4 do not resolve this question because they use different supervision. These omissions make it difficult to assess how much methodological value the proposed recipe adds.

  2. The evidence for broadly effective, recursive improvement is incomplete. In Table 2, three rounds improve SenseNova-SI-Qwen3-VL-8B from 53.54 to 54.78 on average. Approximately 81% of the net gain comes from SPAR-Bench; the other four benchmarks improve by only 0.30 percentage points on average. In Table 1, Gemma and LLaVA gain only about 0.42 points on average, with ViewSpatial regressions of 7.16 and 1.95 points. More importantly, the experiments do not directly validate the proposed mechanism of repeatedly restoring an informative teacher–student advantage. Table 2 omits intermediate-round results for the 8B model, and Figure 4b provides training-set diagnostics rather than tracking teacher and student performance across rounds. The teacher-refresh comparison is restricted to Qwen3-VL-4B/SPAR-Bench. Thus, the results support the usefulness of the recipe in certain settings, but do not establish whether successive rounds reliably produce additional gains or why the supervision remains effective as the model improves. Uncertainty estimates are also needed to interpret the small changes.

  3. Limited methodological novelty relative to prior visual self-distillation. Vision-OPD [1] already uses a privileged visual input—an evidence-centered crop—to construct a teacher from the same model and supervise a full-image student through token-level distillation on student-generated trajectories, without task-answer labels in the distillation objective. It also investigates teacher-update instability and compares fixed, unregularized, trust-region, and EMA teacher strategies (Section 4.3.1). Relative to this prior framework, Spatial-OPSD primarily changes the privileged input to question-specific geometry and adopts round-wise teacher refresh. These differences may be useful, but the paper neither discusses this closely related work nor sufficiently establishes a substantive methodological advance over it, particularly without a comparison against an EMA teacher using the same spatial priors.

[1]:Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation, https://arxiv.org/abs/2605.18740

Questions for authors
  1. Controlled comparison: How does Spatial-OPSD compare with self-SFT using responses generated by the same model with the same spatial priors, without GT answers? Matching the initialization, training data, and training budget would help determine the added value of on-policy distillation.

  2. Clarification and mechanism validation: Could you report the five-benchmark results after each round for SenseNova-SI-Qwen3-VL-8B, with uncertainty estimates, and show whether the privileged teacher retains an accuracy advantage at each round on held-out examples? This would clarify whether the final improvement reflects repeated progress and support the proposed explanation.

  3. Additional generalization evidence: Are the improvements retained on an additional spatial or embodied reasoning evaluation, such as EmbSpatial, the spatial subsets of BLINK, or ERQA? Given the concentration of gains and the observed regressions, such evidence would help assess how broadly the method improves spatial ability.

4 comments

We sincerely thank the reviewer for the careful and insightful assessment and constructive suggestions. We especially appreciate your recognition of the practical value of privileged spatial supervision without inference-time tools, as well as the encouraging results on Qwen3-VL, InternVL, and teacher refresh. Your questions about methodological positioning, Vision-OPD, and the persistence of the teacher–student advantage are valuable and motivated several additional controlled experiments.

1. Contributions beyond the original OPSD

We agree that the paradigm of student on-policy trajectories and dense token-level self-distillation originates from OPSD. Our contributions address two challenges specific to spatial reasoning.

First, the original Self-Distilled Reasoner / OPSD studies mathematical reasoning with ground-truth solutions or verified reasoning traces as privileged teacher context. Spatial-OPSD instead uses question-relevant geometry—depth, instance regions, reconstructed 3D relations, and camera poses—obtained through a perception/reconstruction pipeline and deterministically formatted as spatial context. This transfers reusable scene structure without using task-specific GT answers or answer-derived rewards in the distillation objective.

Second, we introduce round-wise recursive self-improvement: the privileged teacher remains frozen within each round, while the improved student initializes the next round's teacher and student. Reintroducing spatial evidence creates a new teacher–student information asymmetry, supporting repeated improvements. These contributions establish geometry-driven, answer-label-free spatial supervision coupled with recursive model evolution.

2. Relation to Vision-OPD: distinct research challenges and recursive teacher refresh

We thank the reviewer for pointing us to the important Vision-OPD work. We recognize its contribution to label-free visual on-policy distillation, and we will explicitly cite and discuss it in the revision.

The two methods address different starting challenges. Vision-OPD focuses on fine-grained visual perception, enhancing evidence through targeted image crops. Spatial-OPSD addresses more demanding geometric understanding across single images, multiple images and viewpoints, camera motion, and 3D relations. Since these spatial facts can be difficult to recover from visual input alone, our framework integrates a richer tool chain spanning depth, segmentation, 3D reconstruction, geometric relation extraction, and question-relevant spatial subgraphs.

Moreover, we develop recursive self-improvement through a frozen teacher within each round and teacher refresh from the improved student between rounds, rather than applying privileged evidence for a single round.

Following the reviewer's suggestion, we also compared per-step EMA teacher updates (α = 0.05) against within-round frozen teachers using matched geometry, data, and training budgets:

Teacher updateSPARMindCubeMMSIViewSpatialVSIAvg.
Frozen teacher within round44.78831.81029.00040.70055.73040.406
Per-step EMA (α = 0.05)40.46031.04829.10040.21443.64936.894

The within-round frozen teacher outperforms per-step EMA by 3.512 points on the five-benchmark average. This controlled comparison supports stable teacher supervision within a round and recursive updates at round boundaries.

3. Teacher-argmax self-SFT baseline

We appreciate the reviewer's suggestion to include an answer-label-free self-SFT control. The privileged teacher greedily generates pseudo-answers with spatial context, and the student learns from those answers through ordinary SFT. Spatial-OPSD instead distills the privileged teacher's token distribution on the student's own on-policy trajectories.

MethodSPARMindCubeMMSIViewSpatialVSIAvg.
Spatial-OPSD44.78831.81029.00040.70055.73040.406
Teacher-argmax self-SFT43.64833.23828.10039.75851.71439.292

Spatial-OPSD achieves a five-benchmark average of 40.406, above greedy self-SFT at 39.292. This directly addresses the suggested pseudo-answer SFT alternative.

4. Round-wise improvements, held-out teacher advantage, and three training seeds

We thank the reviewer for the insightful questions about continued teacher improvement. We extended the Qwen3-VL-4B round-wise analysis across all five benchmarks:

Strategy / roundSPARMindCubeMMSIViewSpatialVSIAvg.
Shared R144.78831.81029.00040.70055.73040.406
Round-wise R247.46431.14329.50041.82055.94341.174
Round-wise R347.63632.85829.50041.75056.18741.586
Fixed Teacher R244.17033.62028.20040.47253.26839.946
Fixed Teacher R343.16837.04828.50040.35050.00439.814

The round-wise five-benchmark average increases from 40.406 → 41.174 → 41.586, exceeding fixed-teacher R3 by 1.772 points.

On held-out spatial questions, we compared the same model without geometry (student condition) and with privileged geometry (teacher condition):

Diagnostic roundTeacher − Student95% Bootstrap CI
R0+22.865[18.764, 26.826]
R1+14.244[10.848, 17.634]
R2+8.713[5.670, 11.821]
R3+8.050[5.162, 10.914]

The privileged teacher retains a positive advantage across all rounds and remains 8.050 points ahead of the student at R3. Thus, geometric evidence continues to provide informative supervision as the student improves.

We also report results across three independent training seeds on all five benchmarks:

SeedSPARMindCubeMMSIViewSpatialVSIAvg.
4244.78831.81029.00040.70055.73040.406
4342.37234.28628.70040.57756.62940.513
4443.91231.52429.10040.08755.97440.119
Mean ± SD43.691 ± 1.22332.540 ± 1.51928.933 ± 0.20840.455 ± 0.32456.111 ± 0.46540.346 ± 0.203

The five-benchmark average is 40.346 ± 0.203 across the three seeds, demonstrating stable performance.

5. Benefits beyond SPAR training data

To assess dependence on SPAR training examples, we also trained No-SPAR OPSD with SPAR examples excluded and evaluated all five benchmarks:

ConfigurationSPARMindCubeMMSIViewSpatialVSIAvg.
Base35.14824.57028.00039.01055.43036.432
No-SPAR OPSD37.26426.38028.34739.58255.86437.487
Spatial-OPSD44.78831.81029.00040.70055.73040.406

With SPAR training data removed, No-SPAR OPSD still improves on Base across all five benchmarks, raising the average from 36.432 to 37.487 (+1.06 points). This supports applicability beyond a particular training corpus.

6. Additional spatial benchmarks: BLINK and EmbSpatial

We thank the reviewer for suggesting evaluations beyond the original spatial benchmarks. We additionally evaluated Qwen3-VL-4B on BLINK and EmbSpatial:

BenchmarkBaseSpatial-OPSDImprovement (points)
BLINK71.00%71.50%+0.50
EmbSpatial75.577%76.346%+0.769

Spatial-OPSD improves on both BLINK and EmbSpatial. These additional results further demonstrate the transfer of spatial reasoning capabilities beyond the original five benchmarks.

We thank the reviewer again for identifying the most important questions for sharpening this work. In the revision, we will cite Vision-OPD explicitly, distinguish its crop-conditioned visual learning from our multi-view geometric supervision, and clarify the differences from the solution-conditioned original OPSD. The added teacher-argmax self-SFT, EMA, round-wise, held-out, No-SPAR, BLINK, and EmbSpatial results provide further evidence of the method's effectiveness, recursive training mechanism, and applicability to additional spatial evaluations.

Thank you for the detailed rebuttal and the additional experiments. I appreciate the authors' efforts, and the new results provide useful evidence for the effectiveness of the proposed approach.

However, my primary concern regarding methodological novelty remains unresolved. At its core, Spatial-OPSD appears to be a multi-round extension of existing privileged-context OPSD, where the privileged information is replaced with spatial geometry and the teacher is periodically refreshed using the updated student. While these are reasonable design choices, I do not see a fundamentally new learning mechanism beyond repeatedly applying the established OPSD framework.

The comparison with EMA demonstrates that freezing the teacher within each round is empirically beneficial, but this primarily validates a training schedule rather than establishing a substantial methodological contribution. Similarly, the distinction from Vision-OPD appears to lie mainly in the type of privileged information and the application domain.

I acknowledge the additional empirical evidence, including the self-SFT comparison and multi-round improvements. Nevertheless, the improvements over self-SFT and on additional benchmarks remain modest, and the requested intermediate-round results for the stronger 8B model are still missing.

Overall, while I recognize the practical value of applying privileged-context OPSD to spatial reasoning, I remain unconvinced that the proposed method offers sufficient novelty beyond a multi-round adaptation of existing techniques. Therefore, I will maintain my original rating of 2.

We sincerely thank the reviewer for the thoughtful follow-up and for recognizing the usefulness of our additional experiments. We appreciate your continued concern regarding methodological novelty and the request for intermediate-round results on the stronger 8B model.

Following your suggestion, we now provide the complete round-wise results on SenseNova-SI-Qwen3-VL-8B, an already spatially specialized model.

RoundSPARMindCubeMMSIViewSpatialVSIAvg.
R0 (Base)40.873.737.751.264.353.54
R144.373.837.851.364.254.28
R245.574.138.051.464.154.62
R345.874.438.151.464.254.78

The five-benchmark average improves consistently across rounds (53.54 → 54.28 → 54.62 → 54.78), with a total gain of 1.24 percentage points. These results extend our round-wise evidence to a stronger, spatially specialized model.

Regarding methodological novelty, we appreciate the reviewer's perspective and agree that Spatial-OPSD has a relatively simple formulation. This simplicity is intentional. We believe that the simplicity of a method should be distinguished from the novelty of the research problem it addresses and the contribution it makes. Our goal is not to introduce a more complicated distillation objective, but to establish a simple and effective framework for geometry-driven model self-evolution. Rather than relying on verified task-specific answers or a more capable external teacher, Spatial-OPSD transforms reusable scene geometry into privileged supervision and renews the teacher using the improved student between rounds. In this way, the teacher's advantage arises from information asymmetry rather than model capability asymmetry.

The value of this design is supported by our experiments. As reported in Point 4 of our previous rebuttal, round-wise teacher refresh achieves a five-benchmark average of 41.586 on Qwen3-VL-4B, compared with 39.814 for the fixed-teacher strategy, while privileged geometry still provides an 8.050-point held-out teacher advantage at R3. Together with the new 8B round-wise results above, these findings support the effectiveness of our framework for recursive spatial self-improvement. We believe that achieving this capability through a simple and effective design is a strength of Spatial-OPSD. We hope that its contribution can be considered in terms of the research problem it addresses and the learning capability it enables, rather than solely the complexity of its constituent operations.

We sincerely hope that these additional results and clarifications address your remaining concerns. If there are any additional questions and feedback you'd like us to consider, please let us know. Your insights are very valuable to us, and we would be happy to discuss any remaining questions.

Reviewer_d1FbOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(−1)
Read review
Paper summary

This paper presents Spatial-OPSD, a label-free self-improvement framework for enhancing VLM spatial reasoning. During training, a privileged teacher is provided with automatically derived spatial cues, such as depth and camera geometry, while the student receives only the original visual input and question. The teacher then supervises the student on its own generated trajectories through token-level on-policy distillation, which allows spatial knowledge to be internalized without requiring ground-truth labels. The authors further introduce a round-wise training strategy that periodically updates the teacher while keeping it fixed within each round. Experiments across multiple VLM families and five benchmarks demonstrate noticeable overall improvements.

Critical strengths
  • The proposed round-wise recursive self-improvement strategy is effective, which enables the model to continuously benefit from privileged spatial supervision while maintaining training stability.
  • The experiments are comprehensive, covering four different VLM architectures and five spatial reasoning benchmarks, demonstrating the broad applicability of the method.
Critical weaknesses

The paper lacks concrete qualitative examples of both the training inputs and evaluation outputs. In Section 3, it's stated that the spatial context zz, including depth map, instance masks, and 3D relations, is converted into a deterministic text block for the teacher, but the exact representation and prompting format are not illustrated. I would like to see more qualitative examples in the appendix to give a me a better picture of how the OPSD pipeline works. Moreover, qualitative comparisons of model predictions before and after training would help clarify their difference in the reasoning pattern.

Questions for authors
  1. The performance is not consistent across benchmarks (e.g., SPAR-Bench improves well, but see Gemma-3-4B drops from 31.9 to 24.8 on ViewSpatial). Could the authors provide a more fine-grained analysis of which spatial capabilities (e.g., depth estimation, perspective taking) are improved or weakened by Spatial-OPSD? The analysis can be based on qualitative comparisons of cases or performance on benchmark subtasks.

  2. The gains on SenseNova-SI are relatively small, possibly because much of its training data has already been annotated using similar perception and reconstruction tools during QA curation. In this case, how is Spatial-OPSD fundamentally different from directly using these tools to help synthesize spatial QAs with ground-truth answers?

  3. Minor writing suggestion: The presentation of Figures 1 and 2 could be better integrated with the main text. The current discussion does not clearly indicate which textual descriptions correspond to the individual components of each figure. Add \citep links to Figure 1 and 2 in the text in order to make both the figures and the proposed framework easier to understand.

    • The appendix compares Spatial-OPSD with SFT. Unfortunately, I could hardly see the advantage of Spatial-OPSD becuase it outperforms SFT only on SPAR-Bench, which is an in-domain benchmark as SPAR-7M is part of the training data. In other words, I would still prefer using SFT with the off-the-shelf training data containg ground-truth labels as it is more convenient and their performance gains are almost same. So what is the advantage of Spatial-OPSD since all the training data you use come from datasets already containing ground-truth answers?
    • I believe that the practical motivation would be strengthened IF YOU COULD evaluate whether Spatial-OPSD can provide further gains when applied after SFT, or IF YOU COULD illustrate that generating QA supervision with external tools is substantially more costly than constructing the spatial supervision required by Spatial-OPSD.
6 comments

We sincerely thank the reviewer for the encouraging evaluation and thoughtful, constructive suggestions. We particularly appreciate your recognition of the effectiveness and stability of our round-wise recursive self-improvement strategy, as well as the breadth of our experiments across four VLM architectures and five spatial reasoning benchmarks. Your suggestions on concrete examples, fine-grained spatial abilities, and the role of QA supervision are valuable for clarifying our approach. We address them in turn.

1. A concrete cross-view localization example: full prompt, privileged reference, and corrected prediction

We thank the reviewer for suggesting concrete prompts and spatial context. The following two-image object-localization example makes the information available to the student and privileged teacher explicit.

Full question (shared by student and teacher):

<image>
<image>
Find the bounding box of jar in the second image, based on its location in the red bbox of the first image. Select the correct answer from the given choices.
A. [528, 314, 1000, 999]
B. [126, 286, 513, 998]
C. [175, 0, 800, 75]
D. [845, 118, 988, 308]
Your answer can only include one of the options A, B, C, or D.

Additional reference available only to the privileged teacher:

Reference information:
- projected red target center in the second view [x,y] on the 0-1000 image plane: [923.4, 236.2]

Teacher predictions before and after adding the reference:

Teacher input conditionAdditional spatial referenceTeacher predictionGround truthOutcome
Without reference (original inference)NoneADIncorrect
With reference (privileged teacher)Projected center in view 2: [923.4, 236.2]DDCorrect

2. Fine-grained spatial ability analysis

We appreciate the request to look beyond aggregate benchmark scores and identify the spatial abilities most affected by Spatial-OPSD. We decomposed the supplementary Qwen3-VL-4B evaluation by task type. Representative improvements (percentage points) include:

BenchmarkSpatial skill / subtaskImprovement
SPARSingle-view depth estimation (depth_prediction_oc)+30.200
SPARSingle-view distance estimation (distance_prediction_oc)+33.400
SPARCamera motion inference (camera_motion_infer)+22.000
SPARView change inference (view_change_infer)+15.920
MindCubeSpatial relation around+6.800
MindCubeRotation reasoning+3.000
ViewSpatialPerson-perspective object orientation+8.936

These task-level results highlight gains extending from depth and metric distance perception to viewpoint changes, camera motion, and person-centered orientation.

3. Why OPSD rather than using geometry to synthesize QA labels for SFT?

We appreciate this practical question. Spatial-OPSD offers two complementary advantages.

First, Spatial-OPSD outperforms greedy self-SFT using privileged-teacher pseudo-answers. Following the reviewer's suggestion, we added a Teacher-Argmax baseline that generates pseudo-answers using spatial context and trains the student via SFT on the original inputs:

MethodSPARMindCubeMMSIViewSpatialVSIAvg.
Spatial-OPSD44.78831.81029.00040.70055.73040.406
Greedy self-SFT (Teacher Argmax)43.64833.23828.10039.75851.71439.292

Spatial-OPSD achieves a five-benchmark average of 40.406 versus 39.292 for greedy self-SFT, supporting the value of privileged distributional supervision on student-generated trajectories.

Second, geometric priors are not ready-made GT answers, and producing verified QA labels adds cost. Depth, poses, and 3D distances provide reusable scene facts, but multi-view correspondence, hypothetical motion, and compositional spatial questions may require additional reasoning. Only 19.05% of our training questions yielded GT-matching answers under the tested deterministic direct-geometric rules. Constructing reliable QA targets for SFT or GT-derived rewards for GRPO may therefore require stronger VLMs (e.g., Claude), specialized solvers, or human verification, incurring additional inference and annotation costs. Spatial-OPSD instead directly distills spatial evidence without making per-question GT QA labels a prerequisite.

4. Round-wise recursive improvement and clearer Figure 1/2 references

We appreciate your recognition of the recursive teacher-refresh design. We extended the analysis to a complete five-benchmark round-wise/fixed-teacher comparison:

Teacher update / roundSPARMindCubeMMSIViewSpatialVSIAvg.
Shared R144.78831.81029.00040.70055.73040.406
Round-wise R247.46431.14329.50041.82055.94341.174
Round-wise R347.63632.85829.50041.75056.18741.586
Fixed Teacher R244.17033.62028.20040.47253.26839.946
Fixed Teacher R343.16837.04828.50040.35050.00439.814

The round-wise five-benchmark average increases from 40.406 → 41.174 → 41.586; at R3 it exceeds the fixed-teacher variant by 1.772 points.

We also appreciate the clear writing suggestion. We will explicitly link Figure 1 to geometry acquisition and deterministic spatial-text construction, and Figure 2 to student on-policy trajectories, privileged distillation, and round-boundary teacher refresh, alongside the concrete prompt example in the Appendix.

We thank the reviewer again for the positive assessment and constructive guidance. The cross-view prediction case, skill-level analysis, and round-wise results illustrate how Spatial-OPSD transfers reusable geometric evidence into stronger spatial reasoning without relying on per-question ground-truth QA labels.

I appreciate your detailed rebuttal. It has resolved my concerns partially.

However, I am still uncertain about the advantage of Spatial-OPSD over conventional SFT (Question 3 in your rebuttal). Of course your approach is better than self-SFT. What I would like to talk about is the result in Table 4. It is difficult to tell the strength of OPSD compared with gt-SFT in benchmark performance. You have argued about the cost of deriving GT. However, here are my opinions about cost:

  1. GT needs annotation, so do the spatial questions. Your training gives the teacher and student models many spatial questions to learn. The spatial questions should be reasonable so that the question can be solved based on the input observations, and their difficulty needs to be challenging as well. The synthesis of questions also requires annotation.
  2. Many previous works for spatial data synthesis do not require LLM or human annotation. They inherit annotations from datasets such as ScanNet and ScanNet++, and use specialized visual models (e.g., Segmentation, Depth Estimation) for further annotation [1]. The cost is actually little.
  3. The cost you have compared is for data curation. However, as a user, my need is to finetune a small model for myself, and my major concern is the training cost. Those QA datasets with GT are readily available and I can perform simple SFT or GRPO directly. But I need to continuously update an additional teacher model when using your method, which only results in a similar ultimate performance. If you really want to claim the cost for data curation, you should show a pipeline of how to obtain the (v,q)(v, q) from a raw dataset conveniently instead of building directly upon existing QA datasets (e.g., SPAR-7M-RGBD).

[1] Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence

The idea is interesting but I still have questions about its real advantage. And I would like to maintain my score.

Reviewer_fAH4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Spatial-OPSD, an on-policy self-distillation method for spatial reasoning in VLMs. A frozen teacher (the same checkpoint as the student) receives the image, the question, and a text block of question-relevant geometry zz (depth, distances, 3D relations, camera poses) built from dataset metadata or from Grounding DINO, SAM 2, and VGGT. The student sees only the image and question. It minimizes reverse KL to the teacher on its own sampled responses, with no answer cross-entropy or reward. In multi-round training, the round-rr student initializes both roles in round r+1r+1. Training uses 6.2k examples from SPAR-7M, VSI-590K, and SenseNova-SI. One round improves the five-benchmark average for four 4B VLMs. Three rounds on SenseNova-SI-Qwen3-VL-8B reach the highest open-model average (53.54 → 54.78). Ablations cover teacher-refresh frequency, the distillation loss, and noise in the priors.

Critical strengths
  1. Simple, practical idea. Privileged geometry is used only during training, and the deployed student needs no tools. The released code confirms that answer labels never enter the loss (kd_ratio=1.0; the label field is only logged).
  2. Broad and cheap. The same recipe is applied to four model families using 6.2k examples on a single GPU. SPAR-Bench improves consistently (+3.3 to +9.6), and the average does not decrease for any backbone.
  3. Useful teacher-refresh ablation. Refreshing the teacher every step collapses training (entropy → 0, SPAR → 0), and a fixed teacher degrades across rounds. Only round-wise refresh keeps improving, which is a practical lesson for iterated self-distillation.
Critical weaknesses
  1. "Label-free" is largely nominal. In SPAR-7M and SenseNova-SI, the answers are derived from the same 3D annotations used to build zz. Training also uses the direct-answer setting with thinking disabled (Appendix B), so the student's responses are short. The distillation target is therefore essentially a soft pseudo-label from a teacher that can read the answer off zz. On the same data, SFT or GRPO beats Spatial-OPSD on average for 3 of 4 backbones (Table 4).
  2. The gains are mostly in-distribution. Training includes SPAR-7M, and SPAR-Bench supplies most of the gain. For the 8B model, the other four benchmarks change by only +0.7/+0.4/+0.2/−0.1, and it is unclear whether the base row in Table 2 was re-evaluated under the same pipeline. Gemma and LLaVA gain just +0.4 on average, and Gemma drops 7.2 points on ViewSpatial.
  3. Single seed, with visible noise. No variance is reported. In Figure 5, MindCube swings by about 6 points between 5% and 10% noise, and the round-wise ablation covers only SPAR-Bench on one model.
  4. Missing attribution controls. No run gives the teacher zz from a different scene or with corrupted qualitative relations; the noise test perturbs only numeric values. Without such a control, the gains cannot be attributed to geometric knowledge rather than generic self-distillation or answer-format effects. Figure 4b, the only evidence of transfer, is measured on the training set.
Questions for authors
  1. For what fraction of training questions can the answer be read off, or computed deterministically from, zz? Does the student still improve if the teacher receives zz from another scene or with flipped relations? How does Spatial-OPSD compare with SFT on the teacher's argmax answers?
  2. Do the training scenes overlap with the test sets of SPAR-Bench, VSI-Bench, MMSI, or MindCube? What are the results with SPAR-7M excluded from training? Were all baseline rows in Table 2 re-evaluated with your pipeline?
  3. Can you report at least 3 seeds for Tables 1 and 2, the round-wise ablation on all five benchmarks, and the answer-parse failure rates before and after training?
4 comments

We sincerely thank the reviewer for the careful reading and thoughtful, constructive feedback. We appreciate the questions regarding the role of spatial priors, supervision sources, and experimental evaluation, which give us a valuable opportunity to clarify our motivation and provide additional evidence. Spatial-OPSD turns scene-derived geometry into reusable training supervision through on-policy student-trajectory distillation, without requiring ground-truth QA answers for each training example.

1. Geometric priors do not automatically provide low-cost ground-truth QA answers

We appreciate the reviewer's question about answer supervision. Geometric tools can provide reusable scene facts such as depth, camera poses, 3D positions, and distances. Constructing reliable per-question ground-truth (GT) answers for SFT or GT-grounded correctness rewards for GRPO, however, requires additional answer generation and verification.

While low-level distance comparisons may be computed directly, questions involving cross-view object correspondence, hypothetical motion, and compositional spatial relations may require stronger VLM inference (e.g., Claude), task-specific solvers, or human verification. These steps incur additional compute, API, and annotation costs and may introduce label noise.

Our deterministic answer-recovery analysis found that only 19.05% of the training questions yielded GT-matching answers under the tested direct geometric rules. Thus, having spatial priors is not equivalent to having ready-made, verified GT QA labels.

Spatial-OPSD instead directly converts geometry into privileged on-policy distillation supervision, without requiring GT QA answers or GT-derived rewards in its training objective.

2. GT-SFT/GRPO requires answer supervision; Spatial-OPSD outperforms greedy self-SFT

Table 4's SFT uses GT task answers and GRPO uses GT-based correctness rewards; Spatial-OPSD uses neither. Following the reviewer's suggestion, we added Teacher-Argmax self-SFT, where the privileged teacher greedily generates pseudo-answers and the student learns from them using ordinary SFT:

MethodSPARMindCubeMMSIViewSpatialVSIAvg.
Spatial-OPSD44.78831.81029.00040.70055.73040.406
Greedy self-SFT (Teacher Argmax)43.64833.23828.10039.75851.71439.292

Spatial-OPSD achieves a five-benchmark average of 40.406, above greedy self-SFT at 39.292. This supports the value of privileged on-policy distillation over direct pseudo-answer SFT.

3. Correct, scene-relevant spatial priors drive the gains

We evaluated two controls requested by the reviewer: replacing the teacher's prior with a mismatched prior from another example and flipping qualitative spatial relations (e.g., left/right, above/below, front/behind, and closer/farther):

Teacher contextSPARMindCubeMMSIViewSpatialVSIAvg.
Spatial-OPSD (correct prior)44.78831.81029.00040.70055.73040.406
Mismatched prior36.07235.71427.70038.90152.28138.134
Flipped qualitative relations37.52631.23828.20038.49851.30337.353

Spatial-OPSD with the correct prior averages 40.406, above the mismatched-prior (38.134) and flipped-relation (37.353) variants. These results emphasize the importance of accurate, scene-relevant geometric information.

4. Training-data curation and Table 2 sources

We cleaned the training data at the scene level, using only training scenes distinct from the evaluation scenes. Scores for the other models in Table 2 were obtained from the respective official benchmark websites.

5. No-SPAR training: improvements across all five benchmarks without SPAR training data

We also trained No-SPAR OPSD after removing the SPAR training examples and compared it with Base and full Spatial-OPSD:

ConfigurationSPARMindCubeMMSIViewSpatialVSIAvg.
Base35.14824.57028.00039.01055.43036.432
No-SPAR OPSD37.26426.38028.34739.58255.86437.487
Spatial-OPSD44.78831.81029.00040.70055.73040.406

Even without SPAR training data, No-SPAR OPSD improves over Base on all five benchmarks, raising the average from 36.432 to 37.487 (+1.06 points). This demonstrates benefits beyond one specific training corpus.

6. Three training seeds and five-benchmark round-wise results

We evaluated Qwen3-VL-4B across three independent training seeds:

SeedSPARMindCubeMMSIViewSpatialVSIAvg.
4244.78831.81029.00040.70055.73040.406
4342.37234.28628.70040.57756.62940.513
4443.91231.52429.10040.08755.97440.119
Mean ± SD43.691 ± 1.22332.540 ± 1.51928.933 ± 0.20840.455 ± 0.32456.111 ± 0.46540.346 ± 0.203

The five-benchmark mean is 40.346 ± 0.203 across three seeds, demonstrating stable performance. Our five-benchmark round-wise comparison further shows:

Strategy / roundSPARMindCubeMMSIViewSpatialVSIAvg.
Shared R144.78831.81029.00040.70055.73040.406
Round-wise R247.46431.14329.50041.82055.94341.174
Round-wise R347.63632.85829.50041.75056.18741.586
Fixed R244.17033.62028.20040.47253.26839.946
Fixed R343.16837.04828.50040.35050.00439.814

The round-wise five-benchmark average improves from 40.406 (R1) to 41.174 (R2) and 41.586 (R3). R3 improves over R1 by +1.180 points and outperforms fixed-teacher R3 by +1.772 points, demonstrating the benefit of round-wise teacher refresh.

In summary, Spatial-OPSD directly converts scene-derived spatial evidence into answer-label-free learning signals. The new self-SFT, spatial-prior, No-SPAR training, multi-seed, and round-wise experiments provide evidence for the method's effectiveness, applicability beyond SPAR training data, stability, and continued gains from recursive teacher updates.

Reviewer_fAH4OpenReview ↗

Official Comment by Reviewer fAH4

Most of the gains outside SPAR seem within the uncertainty introduced by the seed or survive mismatching of the prior, so gains cannot be attributed to any geometric knowledge. Since the training sets have GT labels, there is no benefit of label-free training in this experiment. Label-supervised SFT/GRPO outperforms label-free training in most backbones.

Most of the gains in Table 2 of the headline correspond to gains from comparing the baseline results to our pipeline. However, most of this gain comes from in-distribution SPAR. To strengthen the paper, I would suggest re-running the base model in the same pipeline and also reporting the accuracy of the teacher model with and without prior.

We thank the reviewer for the thoughtful follow-up. We appreciate the importance of careful attribution and consistent evaluation. To clarify the remaining concerns, we would like to revisit several relevant evaluation details and experimental results already reported in the original submission and our previous rebuttal. For ease of reference, we summarize these findings below.

1. Our Base evaluations already follow the benchmark protocols and use the same configuration as the trained models.

Regarding the suggestion to re-run the Base model in the same pipeline, we would like to clarify that the matched evaluation protocol was already described in Appendix B.1 of the original submission. Our evaluation uses EASI v0.2.2 with its lmms-eval backend and native model adapters, following the benchmark-provided task prompts, answer parsers, and aggregation code without modification. We use zero-shot evaluation and deterministic decoding. These choices align our evaluation procedure with the published benchmark protocols. Specifically, Appendix B.1 states:

All final base-versus-trained comparisons use identical model adapters, processors, decoding settings, and scorers.

Thus, the Base models have already been evaluated under the same benchmark-aligned settings as our trained models; the Base-to-Spatial-OPSD differences do not arise from an evaluation-pipeline change.

2. The non-SPAR evidence goes beyond seed variation and SPAR-dependent training.

As reported in Point 6 of our previous rebuttal, we evaluated three independent training seeds across all five benchmarks. On MindCube, the Base model scores 24.570, whereas Spatial-OPSD achieves 31.810, 34.286, and 31.524 (32.540 ± 1.519). All three trained runs substantially exceed the reported Base result. While the magnitude of improvement varies across benchmarks, these results provide evidence of robust gains beyond SPAR.

Moreover, Point 5 of our previous rebuttal reports a No-SPAR experiment specifically designed to examine dependence on SPAR training data by removing all SPAR training examples. No-SPAR OPSD exceeds Base on all five benchmarks, and its five-benchmark average increases from 36.432 to 37.487 (+1.055 points). This directly provides evidence of improvements without in-distribution SPAR training examples.

3. The prior-intervention and held-out evaluations provide direct evidence that geometry matters.

Following the reviewer's original suggestion, we presented both mismatched-prior and flipped-qualitative-relation interventions in Point 3 of our previous rebuttal. Their five-benchmark averages are 38.134 and 37.353, respectively, compared with the paper’s reported Spatial-OPSD result of 40.406. We appreciate the concern about attributing the gains. However, the persistence of some gains under mismatched priors does not rule out a contribution from correct geometry. We believe the results in Point 3, together with the held-out evidence below, help clarify this distinction.

Regarding the teacher's performance with and without priors, Point 4 of our response to Reviewer CWfr reports a related paired comparison. For ease of reference, we reproduce the paired results for the same model on held-out spatial questions, with and without privileged geometry, across successive rounds:

Diagnostic roundTeacher − Student (points)95% bootstrap CI
R0+22.865[18.764, 26.826]
R1+14.244[10.848, 17.634]
R2+8.713[5.670, 11.821]
R3+8.050[5.162, 10.914]

The geometry-conditioned model achieves a positive advantage in every round, including +8.050 points at R3 with a 95% bootstrap confidence interval of [5.162, 10.914]. This is direct evidence of the utility of privileged geometry on held-out questions. The table reports paired accuracy differences, rather than separate absolute accuracies.

We sincerely thank the reviewer for the continued discussion. We sincerely hope that these clarifications address your remaining concerns. If there are any additional questions and feedback you'd like us to consider, please let us know. Your insights are very valuable to us, and we would be happy to discuss any remaining questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.