acceptodds
Under review as a conference paper at ICLR 2027

SpatialOPSD: Self-Distilling Spatial Intelligence from Verified Coding-Agent Traces

Abstract

Spatial coding agents significantly improve spatial reasoning in Multimodal Large Language Models (MLLMs) by using external tools to generate verified execution traces. However, this paradigm inherently suffers from prohibitive inference-time overhead and external dependencies. In this paper, we explore whether an MLLM can internalize this agentic capability to operate entirely tool-free. We begin with a simple observation: prompting an MLLM with summarized execution traces of a spatial coding agent naturally unlocks the model’s internal spatial Chain-of-Thought (CoT). Motivated by this, we introduce SpatialOPSD, an on-policy self-distillation framework that internalizes spatial reasoning into a standalone MLLM by formulating verified agent traces as privileged information. To mitigate privileged-information leakage during distillation, we introduce Repetition-Aware Distillation, which combines repetition masking with unlikelihood regularization. Experiments across multiple benchmarks demonstrate that self-distilling SpatialOPSD achieves higher average accuracy than SFT and GRPO on both spatial and OOD datasets, exhibiting superior performance and generalization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Coding Agent, Spatial Intelligence, On-policy self-distillation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dRs3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SpatialOPSD uses a frozen model conditioned on successful coding-agent traces to train a student that reasons without tools. It adds repetition masking and an unlikelihood penalty, then evaluates spatial reasoning and transfer.

Critical strengths

The ViewSpatial gains are useful evidence of transfer beyond the training source. Results across model sizes and the privileged-context ablations also give the approach more support than a single benchmark result would. The practical goal is clear: use expensive agent traces during training while keeping inference free of tool calls. The experiments address that setting directly, and the frozen teacher makes the source of privileged context reasonably easy to follow.

Critical weaknesses

The leakage metric is my main concern. Appendix A.3 counts any use of “agent”, “trace”, or “tool”, including ordinary uses. A lower score could mean a change in vocabulary rather than less reliance on unavailable information. Table 3 also favors masking alone on accuracy, at 53.0% versus 51.9% for the full method. Unlikelihood alone has lower measured leakage. These variants make different tradeoffs, but the lexical metric makes those tradeoffs hard to interpret. A semantic check is needed before treating the lower keyword rate as a correction of the underlying failure.

The SFT baseline uses 4,459 tool-free rationales, while OPSD uses 5,008 agent-correct instances. The comparison changes the training examples and targets as well as the objective. This leaves open whether OPSD improves how the model learns from traces, or whether the benefit comes mainly from which traces it sees.

Questions for authors

Can you manually check a sample of keyword matches for actual unsupported trace references? How does SFT on the same summarized agent traces compare with OPSD? The matched SFT comparison would help distinguish the value of agent traces from the value of the OPSD objective.

4 comments
AuthorsOpenReview ↗

Response to Reviewer dRs3

We deeply thank Reviewer dRs3 for the insightful review! We greatly appreciate the reviewer's recognition of our practical tool-free inference goal, clear experimental setup, and strong transfer results on ViewSpatial. Directly following Reviewer dRs3’s valuable suggestions, we conducted targeted experiments to fully resolve the two main concerns: (1) We implemented a manual audit and LLM-based semantic check as suggested, verifying that SpatialOPSD massively reduces genuine semantic reliance on unavailable traces (from 82.5% to 25.1%), not just keyword usage; and (2) We evaluated SFT and SpatialOPSD on the exact same 2,644 examples, conclusively showing that our gains (+1.6% average, +8.5% on ViewSpatial) come from the OPSD objective itself rather than data differences. Details are provided below.

Q1: The keyword-based leakage metric might only reflect a vocabulary shift rather than reduced reliance on unavailable traces, making ablation tradeoffs hard to interpret. Can you perform a semantic evaluation and manually check a sample of keyword matches?

A1: We thank the reviewer for this valuable suggestion. Following it, we added both a manual audit and a semantic leakage evaluation.

Manual audit. We manually inspected all 95 responses from SpatialOPSD on MindCube-Tiny that the keyword rule flags. 86 of them (90.5%) contain genuine references to unavailable privileged information, e.g., "The reasoning trace says: > The camera moved diagonally forward and right." The remaining 9 use the keywords in their ordinary sense, either as objects in the scene ("a black tool cabinet") or as the verb trace ("if you trace the line of sight").

Semantic evaluation. We further used Qwen3.7-plus to judge every response. A response counts as leaking if it refers to, quotes, or relies on any source that is absent from the model's input. This includes paraphrased references that contain none of the keywords (e.g., "the provided reasoning"). Ordinary uses of the keywords, quotations of the question, and image descriptions are excluded.

MethodMindCube-TinyViewSpatialRegex-Based Leakage RateSemantic-Based Leakage Rate
Vanilla OPSD55.446.081.682.5
SpatialOPSD56.946.88.325.1

For vanilla OPSD, the keyword and semantic rates nearly coincide, which confirms that the baseline failure is genuine reliance on unavailable traces. Because the semantic criterion is stricter and also captures paraphrased references, it yields a higher absolute rate for SpatialOPSD. Even under this criterion, SpatialOPSD reduces leakage from 82.5% to 25.1%, while also improving accuracy on both benchmarks. This shows that the improvement reflects substantially less reliance on unavailable information rather than a change in vocabulary alone. We will add the judge prompt, and the manual audit to Appendix.

Q2:How does SFT on the same summarized agent traces compare with OPSD? The matched SFT comparison would help distinguish the value of agent traces from the value of the OPSD objective.

A2: We sincerely thank the reviewer for this insightful comment. Raising the question of data parity is crucial for a rigorous evaluation. Does SpatialOPSD’s advantage merely stem from data size or example selection? — No. A strictly controlled experiment on an identical dataset confirms its inherent algorithmic superiority.

To thoroughly address your concern regarding training data discrepancies (originally 4,459 for SFT vs. 5,008 for SpatialOPSD), we conducted an additional ablation study. We extracted the exact intersection of both datasets—yielding 2,644 shared examples—and trained Qwen3.5-9B from scratch using both methods under their respective optimization procedures. This strictly eliminates any confounding factors related to data volume or sample selection.

MethodTrain dataMindCube-TinyViewSpatialAverage
Qwen3.5-9B—48.343.445.9
SFT445958.137.047.6
SpatialOPSD500856.946.851.9
SFT264459.238.048.6
SpatialOPSD264453.946.550.2

Key Takeaways from the Matched-Data Comparison:

  • Consistent Overall Advantage: Even when restricted to the exact same 2,644 examples, SpatialOPSD outperforms SFT on average (50.2% vs. 48.6%, +1.6%), driven by a massive +8.5% surge on ViewSpatial (46.5% vs. 38.0%).

  • Reproduced Qualitative Pattern: The results perfectly mirror our full-data findings—while SFT leans towards MindCube-Tiny (59.2% vs. 53.9%), SpatialOPSD dominates on ViewSpatial and secures a higher overall average.

Conclusion: By strictly locking the training data, this experiment decisively proves that SpatialOPSD’s superior performance is not an artifact of data scale or curation, but a direct result of its distinct supervision and optimization paradigm.

Dear Reviewers,

Please read the authors’ rebuttal and join the discussion. Kindly clarify whether your main concerns have been addressed and raise any follow-up questions promptly so the authors have time to respond.

Thank you for your time and effort!

AuthorsOpenReview ↗

Follow-up before the discussion deadline – Paper 7315

Dear Reviewer dRs3,

Thank you for your time and thoughtful consideration of our submission.

As the discussion period is drawing to a close, we would greatly appreciate your response at your earliest convenience, ideally before the deadline, so that we have an opportunity to address any remaining questions or concerns.

We understand that you have a busy schedule and sincerely appreciate your time and engagement.

Best regards, The authors

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer dRs3,

Thank you for your careful review and thoughtful feedback.

As the discussion deadline approaches, we would appreciate your feedback on whether our responses have addressed your concerns. If any issues remain unresolved, we would be happy to provide further clarification.

Thank you again for your time and consideration.

Best regards, The authors

Reviewer_8Ps4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents SpatialOPSD, a framework for transferring the spatial reasoning of a tool-using multimodal coding agent into a multimodal LLM that answers without tools. It builds on the author's observation that conditioning the model's chain-of-thought (CoT) on a spatial coding agent's execution trace roughly matches the agent's own accuracy. During training, a spatial coding agent, which writes and runs Python code with 3D perception tools, solves the training tasks, with the summary of corrext solutions summarized as privileged information for the teacher during OPSD. To faciltate the traiing the authors propose a Repetition-Aware Distillation loss that removes spans that copy the privileged text or repeat earlier output from the distillation loss, and it penalizes them with an unlikelihood (UL) term. The authors show that SpatialOPSD reaches higher average accuracy on MindCube and ViewSpatial than supervised fine-tuning (SFT) and reinforcement learning (GRPO) baselines, and narrows the gap to the tool-using agent. They also show that it largely preserves out-of-distribution (OOD) performance and improves models of several sizes.

Critical strengths
  1. Internalizing tool use into a model that needs no tools at inference is a useful direction, and it works here: the model beats both trained baselines and shows clear OOD generalization.
  2. I really liked the ide of Repetition-Aware Distillation as a a cheap fix for a real failure mode. While the paper already shows its use in this paper - it attempts to solve a braoder issue with OPSD and I believe can be useful in other domains as well.
Critical weaknesses
  1. Claims are stronger than the per-benchmark evidence. The introduction says SpatialOPSD "substantially surpasses standard SFT and RL baselines" but SFT is higher on MindCube-Tiny (58.1 vs. 56.9, Table 1). All results are single runs without seeds or confidence intervals. In addition, λrep, Lmin, and the SUMMARY format were chosen on the same two test benchmarks (Tables 4 and 6). The average gains still hold, so this does not change my recommendation, but the wording should be softened and some measure of variance reported.
  2. Missing positioning against Visual Program Distillation (VPD) [1]. VPD (CVPR 2024) already executes and verifies tool-calling visual programs. It turns the correct ones into language rationales and distills them into a vision-language model that answers "with a single forward pass", motivated by the cost of running the tools. The paper does not cite it, so contribution 2 ("distills tool-augmented spatial reasoning into multimodal models", p. 2) reads broader than it is. Notably, relative to VPD, three things are new: the on-policy objective with the trace in the teacher's context, the multi-turn spatial-agent traces, and Repetition-Aware Distillation - so the paper still is interesting but will beenfit from a more conservative framing.

References [1] Y. Hu et al. Visual Program Distillation: Distilling Tools and Programmatic Reasoning into Vision-Language Models. CVPR 2024. https://arxiv.org/abs/2312.03052

Questions for authors

Q1. Eq. 7 masks a span if it copies the privileged information or repeats earlier output, but Table 3 ablates Mask vs. UL without separating these two flag types. Could you report accuracy and leakage with Mask and UL applied only to privileged-copy spans and only to self-repetition spans? What share of flagged tokens is of each type?

Q2. In Table 4, raising λrep from 0 to 0.15 lowers MindCube-Tiny accuracy from 60.0 to 55.1, while leakage falls from 73.6% to 0.8%. Mask alone also has the best average in Table 3 (53.0 vs. 51.9). Is this cost driven by one mask type, for example by penalizing legitimate restatements of objects and views? Do the two mask types have competing effects?

Q3. Are the numbers in Tables 1, 3, and 4 from single runs? Could you add seeds or bootstrap confidence intervals? Was a validation split used to choose λrep, Lmin, and the SUMMARY format?

Suggestion

  • It would be interesting to plug SpatialOPSD-9B back into SpatialClaw as the agent's backbone. If the agent then gains less from its tools, the model has internalized what the tools add. If it gains as much as before, the improvement may come more from better use of the visual context the model already has.
7 comments
AuthorsOpenReview ↗

Response to Reviewer 8Ps4 (Part 1/3)

We sincerely thank Reviewer 8Ps4 for the thorough, constructive review and the positive support for our work. We are especially encouraged by your recognition that our overall average gains hold and that our paper remains interesting with three distinct innovations (i.e., the on-policy objective with the trace in the teacher's context, the multi-turn spatial-agent traces, and Repetition-Aware Distillation). We also greatly appreciate your insightful suggestions on softening our claims, properly positioning our work against VPD [1], and conducting finer-grained ablations on the masking mechanisms. We have fully incorporated this feedback and present our detailed responses and additional experimental results below.

W1: Overstated claims & evaluation rigor: Wording in the Intro should be softened since SFT outperforms SpatialOPSD on MindCube-Tiny (58.1 vs. 56.9) and multi-seed variances are missing.

A1: We sincerely thank the reviewer for the thorough reading and constructive suggestions, which have significantly improved the rigor and positioning of our paper. We have incorporated all suggested revisions into the updated manuscript:

  • Refined Wording: In the revised Introduction, we have toned down our claims to precisely state that SpatialOPSD achieves overall average improvements over SFT and RL baselines, while explicitly acknowledging SFT's competitive performance on the MindCube-Tiny subset.
  • 5-Run Mean and Variance: We respectfully clarify that our reported metrics were actually the mean of 5 independent runs, rather than single runs. We apologize for not making this explicit earlier. In the revised manuscript, we have clearly stated this evaluation protocol and added the standard deviations across the 5 runs directly into the main tables (Table 1) to fully address this concern.

W2: The paper lacks a citation to Visual Program Distillation (CVPR 2024). Contribution 2 should be framed more conservatively to highlight the specific novelties over VPD (on-policy objective, multi-turn spatial traces, and Repetition-Aware Distillation).

A2: We sincerely thank the reviewer for the thorough reading and constructive suggestions, which have significantly improved the rigor and positioning of our paper. We have incorporated all suggested revisions into the updated manuscript:

  • Citation and Comparison: We thank the reviewer for pointing out Visual Program Distillation (VPD) and insightfully summarizing our key differences. We have added citations and a detailed discussion of VPD in both the Introduction and Related Work.

  • More Conservative Framing: Following your advice, we have narrowed the scope of Contribution 2. Rather than claiming broad tool-augmented distillation, we now explicitly position SpatialOPSD relative to VPD by highlighting our three distinct technical novelties: (i) the trace-conditioned on-policy distillation objective, (ii) the utilization of multi-turn spatial-agent traces, and (iii) Repetition-Aware Distillation to mitigate degeneration.

AuthorsOpenReview ↗

Response to Reviewer 8Ps4 (Part 2/3)

Q1: Eq. 7 masks a span if it copies the privileged information or repeats earlier output, but Table 3 ablates Mask vs. UL without separating these two flag types. Could you report accuracy and leakage with Mask and UL applied only to privileged-copy spans and only to self-repetition spans? What share of flagged tokens is of each type?

A1: Thank you for this insightful question. To isolate the contributions of each detection mechanism, we conducted additional ablations where both KL masking and the unlikelihood (UL) penalty are triggered exclusively by self-repetition spans (self-only), exclusively by privileged-information copy spans (leakage-only), or by both (the full SpatialOPSD). Note that self-only here modifies the detection source for both mechanisms, distinguishing it from the mechanism-level mask-only ablation in Table 3.

MethodMindCube-Tiny (Acc ↑)ViewSpatial (Acc ↑)PI Leakage (↓)
Vanilla OPSD55.4 ± 0.846.0 ± 0.281.6 ± 0.7
SpatialOPSD (self-only)56.6 ± 0.746.2 ± 0.24.0 ± 0.7
SpatialOPSD (leakage-only)56.8 ± 0.446.0 ± 0.269.4 ± 0.9
SpatialOPSD (full)56.9 ± 0.446.8 ± 0.28.3 ± 0.6

The results show non-additive effects: self-only achieves the lowest leakage, while the combined method achieves the highest mean accuracy on both benchmarks. This is consistent with the normalization in our objective, $ \ell_{\mathrm{rep}}=\frac{1}{M_y}\sum_t m_t[-\log(1-p_t)]. $

For a fixed rollout, adding transcription detections can increase MyM_y, reducing the UL weight on existing self-repetition tokens. The combined mask also changes which positions receive KL supervision. Thus, combining the rules redistributes supervision and regularization; it does not simply add a stronger penalty. This provides a plausible explanation for self-only’s lower leakage, while broader coverage of both failure modes may benefit accuracy. We acknowledge that the full method does not maximize accuracy or minimize leakage individually. The higher $\lambda_{\mathrm{rep}}$ results indicate a trade-off between accuracy and leakage suppression. Our length threshold $L_{\min}=32$ limits penalties on isolated object or view mentions, but does not exclude legitimate longer restatements. These ablations therefore clarify the contribution of each detection source, while not establishing that penalizing legitimate restatements causes the accuracy drop at λrep=0.15\lambda_{\mathrm{rep}}=0.15.

Q2: In Table 4, raising λrep from 0 to 0.15 lowers MindCube-Tiny accuracy from 60.0 to 55.1, while leakage falls from 73.6% to 0.8%. Mask alone also has the best average in Table 3 (53.0 vs. 51.9). Is this cost driven by one mask type, for example by penalizing legitimate restatements of objects and views? Do the two mask types have competing effects?

A2: Thank you for highlighting the accuracy–leakage trade-off. To clarify the contribution of each detection source, we analyzed the composition of masked tokens throughout the full SpatialOPSD training. On average, self-repetition accounts for 92% of all masked tokens, whereas privileged-information (PI) transcription constitutes the remaining minority. While the regularizer primarily penalizes repeated content by volume, the dynamic ratio curve—which will be added to the Appendix—confirms that both mask types actively function throughout training. Note that while this diagnostic reveals where the penalty is concentrated, token proportions alone do not definitively establish which source drives the accuracy fluctuation.

Q3: Are the numbers in Tables 1, 3, and 4 from single runs? Could you add seeds or bootstrap confidence intervals? Was a validation split used to choose λrep, Lmin, and the SUMMARY format?

A3: All reported results are averaged over 5 independent runs. For evaluation, we directly use the checkpoint from the final step of the single training epoch, without employing a separate validation set for hyperparameter tuning or early stopping. During training, no global random seed was explicitly fixed (allowing for natural stochasticity across runs), while the vLLM engine was initialized with its library-default seed (0).

AuthorsOpenReview ↗

Response to Reviewer 8Ps4 (Part 3/3)

Suggestion. It would be interesting to plug SpatialOPSD-9B back into SpatialClaw as the agent's backbone. If the agent then gains less from its tools, the model has internalized what the tools add. If it gains as much as before, the improvement may come more from better use of the visual context the model already has.

We thank the reviewer for suggesting this insightful diagnostic experiment! Following your advice, we plugged SpatialOPSD-9B back into the SpatialClaw agent framework as its backbone and compared it against the original SpatialClaw (powered by the base Qwen3.5-9B). The results are summarized in the table below:

ModelAverageMC: AmongMC: AroundMC: RotationMC: OverallVS: CP-RDVS: CP-OVOVS: PP-OVOVS: PP-RDVS: PP-SS-RDVS: Overall
Q-9B57.357.776.461.562.960.336.750.763.542.651.6
S-9B58.458.283.662.565.161.435.753.458.444.051.7

(Note: Q-9B = Qwen3.5-9B, S-9B = SpatialOPSD-9B; MC = MindCube-Tiny, VS = ViewSpatial)

Adding inference-time tools boosts the base model by +11.4% (45.9% → 57.3%), whereas SpatialOPSD-9B gains a more modest +6.5% (51.9% → 58.4%). Notably, both models reach ~51.7% on ViewSpatial with tools. This indicates that SpatialOPSD has already absorbed most tool-derived gains into its weights. Overall, SpatialOPSD not only strengthens standalone, tool-free spatial reasoning, but also serves as a more capable backbone for tool-augmented agents. We have included this analysis in the revised manuscript.

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

This paper proposes SpatialOPSD, an on-policy self-distillation framework designed to internalize tool-assisted spatial reasoning capabilities into MLLMs. The central idea is to leverage verified execution traces generated by spatial coding agents as privileged information for a frozen teacher model, while training a student model that only observes the original multimodal inputs. Experiments on MindCube-Tiny and ViewSpatial demonstrate that SpatialOPSD improves the average accuracy of Qwen3.5-9B from 45.9% to 51.9%, outperforming the reported SFT and GRPO baselines.

Critical strengths

The paper addresses an important challenge of reducing the inference-time overhead of tool-augmented spatial reasoning. SpatialOPSD presents a simple and effective framework for transferring spatial reasoning capabilities from verified agent traces to MLLMs. The experiments demonstrate consistent improvements in average spatial reasoning accuracy over SFT and GRPO baselines, supported by extensive ablation studies across privileged information types and training configurations.

Critical weaknesses
  1. SpatialOPSD introduces additional computational overhead during training due to agent rollouts, trace verification, summarization, and teacher–student distillation. However, the paper lacks a comprehensive cost analysis to justify whether these additional costs are offset by the inference-time efficiency gains.

  2. Verifying final answers does not guarantee the correctness of reasoning traces. Moreover, distilling teacher predictions does not necessarily imply that the student has acquired spatial reasoning capabilities rather than merely imitating the teacher's output patterns.

  3. Although SpatialOPSD achieves performance improvements, the current results provide limited evidence that distillation offers a distinct or generalizable advantage for enhancing spatial reasoning.

Questions for authors

Refer to the Weaknesses section.

4 comments
AuthorsOpenReview ↗

Response to Reviewer YDKD (Part 1/2)

We sincerely thank Reviewer YDKD for thorough and constructive review. We are encouraged that you found SpatialOPSD to be a simple and effective framework for tackling inference-time overhead, supported by consistent accuracy gains and extensive ablations. The remaining questions you raised are very valuable for strengthening our paper. Below, we fully address all three concerns through newly added cost analyses, theoretical justifications, and empirical proofs:

  • SFT-Level Training Cost (W1): On 8×H100 GPUs, our full pipeline—spanning agent rollout, summarization, and distillation—completes in just 6.23h, closely matching standard SFT (5.72h) while outperforming GRPO (6.74h).

  • Beyond Pattern Imitation (W2): SpatialOPSD shifts learning dynamics from rote memorization to genuine policy internalization by coupling deterministic Python execution feedback with on-policy, anti-copying distillation over student rollouts.

  • Broad Generalizability (W3): Our framework not only surpasses SFT (+4.3%) and RL (+3.7%) without inference-time tools, but also demonstrates robust OOD transfer, consistent scaling (2B–27B), and teacher-pipeline invariance.

We hope these compelling findings resolve your reservations and look forward to engaging with your thoughts on the detailed responses below.

W1: SpatialOPSD introduces additional computational overhead during training due to agent rollouts, trace verification, summarization, and teacher–student distillation. However, the paper lacks a comprehensive cost analysis to justify whether these additional costs are offset by the inference-time efficiency gains.

A1: We thank the reviewer for raising the concern regarding computational overhead. To address this, we tracked the exact end-to-end wall-clock time across all stages of SFT, GRPO, and SpatialOPSD under identical hardware (8 × NVIDIA H100 GPUs) and concurrency settings (64-way API concurrency). All training configurations strictly follow the original settings reported in the Appendix, and the complete training code is provided in the supplementary material.

MethodStage 1: Agent Rollout & Verification (9B)Stage 2: Processing ([Qwen3.7-Plus], 64-way)Stage 3: Formal GPU TrainingTotal Pipeline Wall-Clock
SFT-4.10 h (CoT Generation)1.62 h5.72 h
GRPO--6.74 h (Online Rollouts)6.74 h
SpatialOPSD2.75 h1.09 h (Trace Summary)2.39 h6.23 h

As demonstrated by the profiling results, the concern regarding extra computational burden does not hold in practice, supported by three key findings:

  • Comparable End-to-End Pipeline Time: The total wall-clock time of SpatialOPSD (6.23 h) is on the exact same order of magnitude as both SFT (5.72 h) and GRPO (6.74 h), showing negligible difference in overall time investment.
  • Agent Rollout + Summarization is Faster than Direct Generation (Stages 1 & 2): In our data curation phase, pairing a lightweight 9B agent for trace generation (2.75 h) with Qwen3.7-Plus for trace summarization (1.09 h) takes 3.84 h in total. Because summarizing existing traces is much faster than generating long reasoning chains from scratch, this combined pipeline actually takes less time than prompting Qwen3.7-Plus to directly solve the problems (4.10 h in SFT).
  • SFT-Level Training Efficiency (Stage 3): By decoupling rollouts from the training loop, SpatialOPSD completely avoids GRPO's online generation bottleneck (6.74 h). Consequently, our formal GPU training time (2.39 h) remains highly efficient and closely matches that of standard SFT (1.62 h). In summary, SpatialOPSD introduces no meaningful computational overhead compared to standard baselines, while delivering superior downstream performance. We will incorporate this detailed breakdown into the revised manuscript.
AuthorsOpenReview ↗

Response to Reviewer YDKD (Part 2/2)

W2: Final-answer checks do not ensure valid traces, and distillation may cause pattern imitation rather than true spatial reasoning.

A2: We fully agree with this insightful observation. In fact, Preventing "hallucinated traces" and "superficial imitation" is the exact motivation behind SpatialOPSD, which we address via two core mechanisms:

  1. Execution-Grounded Supervision: Unlike free-form text where flawed logic can accidentally yield right answers, our traces are strictly anchored in a live Python environment. Every intermediate step must execute as valid code, producing deterministic tool outputs or runtime errors that constrain subsequent steps. Combining outcome verification with executable code constraints acts as a rigorous filter against spurious reasoning.

  2. On-Policy Distillation & Anti-Copying: Instead of passively memorizing fixed teacher templates as in SFT, SpatialOPSD prevents mere pattern imitation by combining student-driven rollouts—where the student actively generates its own prefixes from visual inputs and receives dynamic token-level teacher guidance within its own policy distribution—with explicit anti-copying, where our Repetition-Aware Distillation masks copied/repetitive spans and applies an unlikelihood penalty to suppress rote memorization.

To vividly illustrate how these mechanisms prevent spurious traces and mere imitation in practice, we provide two representative examples below:

SFT Teacher:

To solve this, let's break down the spatial relationships... 
1. Analyze Image 1: This image shows a view looking straight out a window... 
2. Analyze the Camera Movement: ...taken by turning the camera 90 degrees to the right (clockwise)... 
3. Analyze Image 2:...shows a white armchair with a yellow cushion, a lamp... 
4. Simulate the Person's Movement: ...Turn 90 degrees to the left... 
5. Identify the Object: ...Therefore, after turning 90 degrees to the left, the sofa with the potted flowers is to your right

SFT Student:

To solve this, let's break down the spatial relationships... 
1. Analyze Image 1: This image shows a view looking out a window... 
2. Analyze the Camera Movement: ...taken by turning the camera 90 degrees to the right (clockwise)... 
3. Analyze Image 2: ...shows a white armchair with a yellow cushion, a lamp... 
4. Determine the Final Orientation: ...Turn 90 degrees left... 
5. Identify What is to the Right: ...Therefore, after turning 90 degrees to the left from the initial position, the sofa with the potted flowers is to your right.

SFT student blindly regurgitates the teacher's exact 5-step template and phrasing almost word-for-word, showing pure pattern memorization.

SpatialOPSD Teacher (Privileged):

Let’s analyze... provided in the agent trace... In image 2, the camera is positioned such that... the agent trace computes distance to 'one of the single sofas' and shows it decreases... In image 2 BEV plot (in trace), Camera 1 is at bottom center... \boxed{B}

SpatialOPSD Student (Tool-Free):

Let’s carefully analyze the situation using the provided images and the question. 
Step 1: Viewpoint in Image 2: A white basket in the foreground; left is a beige sofa... 
Step 2: What does 'turn right' mean? The two single sofas (armchairs in Image 3 & 4) are off-camera to the right in Image 2... 
Step 3: Spatial reasoning: Turning right from Image 2 aligns with Image 3’s viewpoint, walking straight toward the armchairs... \boxed{B}

SpatialOPSD student avoids copying any trace/BEV jargon, formulating an independent, visually grounded 4-step geometric proof.

W3: Although SpatialOPSD achieves performance improvements, the current results provide limited evidence that distillation offers a distinct or generalizable advantage for enhancing spatial reasoning.

A3: We respectfully point out that our paper already provides extensive empirical evidence proving both the distinct advantage of SpatialOPSD over other paradigms and its generalizability across domains, model scales, and agent pipelines:

  1. Distinct Advantage over SFT & RL: Unlike standard paradigms, SpatialOPSD uniquely uses verified execution traces as privileged teacher context to guide student rollouts while penalizing rote copying. In Table 1, this boosts average spatial accuracy from 45.9% to 51.9%, substantially outperforming both SFT (+4.3%) and GRPO (+3.7%).

  2. Broad Generalizability: Trained solely on MindCube, SpatialOPSD transfers zero-shot to unseen spatial benchmarks (Table 2)—improving OmniSpatial (53.7%→54.2%) and BLINK (67.8%→68.9%) while preserving general capabilities on MMStar and Video-MME. Moreover, these gains hold consistently across model scales (Table 5) and remain agnostic to the teacher pipeline (Table 8), achieving equally strong results when distilling from two completely different agents.

AuthorsOpenReview ↗

Follow-up before the discussion deadline – Paper 7315

Dear Reviewer YDKD,

Thank you for your careful review. We hope that our responses have addressed your concerns and clarified the points raised in your review.

As the discussion deadline is approaching, we would greatly appreciate your feedback on our responses and any concerns that remain unresolved. This would allow us to provide further clarification and help the Area Chair make a well-informed and balanced assessment of our work.

Thank you again for your time and consideration.

Best regards, The authors

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer YDKD,

As the discussion deadline approaches, we would appreciate your feedback on whether our responses have addressed your concerns. If any issues remain unresolved, we would be happy to provide further clarification. Thank you again for your time and consideration.

Best regards, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.