acceptodds
Under review as a conference paper at ICLR 2027

Beyond Spatial Benchmarks: From Spatial Reasoning to Navigation

12 views

Abstract

Does progress on spatial reasoning benchmarks translate into better navigation? Existing benchmarks test isolated inferences from images or videos, with little connection to downstream navigation. Our analysis reveals a gap between benchmark-oriented spatial specialization and navigation performance, and shows how aligning spatial supervision with navigation goals, phases, and decision learning improves navigation. Guided by these findings, we build Spatial-Nav-100K and fine-tune in two stages, i.e. first learning a shared spatial-navigation foundation, and then specializing each phase with the abilities it relies on. We further introduce Spatial-NPD, where a teacher conditioned on spatial priors produces grounded action preferences and distills them into a student policy, so no explicit spatial reasoning is needed at inference. With 45 A100 GPU-hours of policy training, our 8B model reaches SR/SPL of 77.4/35.4 on HM3D-v0.2, 60.2/30.5 on HM3D-v0.1, and 47.9/20.6 on train-unseen MP3D. It outperforms several systems that rely on closed-source models or thousands of GPU-hours of training, at 148 ms per action step. All code and datasets will be publicly available.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied Navigation, Spatial Reasoning, Supervised Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zYwVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether spatial reasoning improvements of VLMs can be translated into better navigation capability. Based on experiments, it finds out that the spatially specialized VLMs perform worse than their base models when inserted into existing navigation systems and spatial supervision becomes useful when it is connected to the navigation goal. To this end, it proposes a spatial-nav-100K dataset and use it for two-stage fine-tuning and distillation. As illustrated in the results, the final distilled model Spatial-Nav achieves strong performance compared to existing methods.

Critical strengths

The idea of connecting spatial reasoning question with navigation capability is interesting and based on the extensive experiments and analysis, the authors do show fine-tuning on spatial reasoning question can improve model's capability on navigation. Based on the results, although the proposed model does not achieve state-of-the-art performance, it does show comparable capability and as a illustration of the effectiveness of the paper's idea.

Critical weaknesses
  1. My major concern stems from the distillation process. While the author mentions the two-stage finetuning improve the spatial intelligence of the model but leave the link from spatial evidence to decisions implicit, I think when you distill the model with a focus on the link from spatial to action, you will inevitably damage the spatial reasoning capability compared to the previous fine-tuned model. Could the author provides the results of teacher model on spatial-nav-100K and the results of distilled model on spatial reasoning benchmarks such as HM3D and VSI-Bench.
  2. I'm curious about the exact meaning, evaluation and samples for ability types proposed such as visibility, topology and affordance. Could the author elaborate more on this part and also give more details about these samples inside their proposed benchmark.
Questions for authors

Similar as weaknesses.

2 comments

Response to Reviewer zYwV

Conclusion

Thank you for the review. We evaluated each phase expert on six spatial benchmarks after Spatial-NPD. Overall, distillation indeed shifts the model's focus toward translating spatial understanding into actions, but the resulting degradation in its general spatial capabilities remains relatively modest.

In addition, we have also provided more details of the proposed Spatial-Nav-100K dataset and released the anonymized code repository and dataset.

W1. Does action-oriented distillation damage spatial reasoning?

Teacher on Spatial-Nav-100K. Supplementary Table 6 reports the teacher on the Spatial-Nav-100K development split. Action accuracy is 83.03% without a prior (Student), 88.02% with the aligned spatial prior (Teacher).

ModelAction Acc. ↑Action NLL ↓
Navigation context xtx_t only83.030.4541
xtx_t + spatial prior88.020.3439

Distilled model on HM3D. HM3D is a navigation benchmark, so we report the navigation metric SR / SPL here. Relative to training without Spatial-NPD, the distilled model improves SR / SPL by +2.9 / +2.5 on HM3D-v0.2, +2.6 / +1.3 on HM3D-v0.1, and +0.8 / +1.3 on MP3D (Main Table 7a).

Distilled model on spatial benchmarks. Supplementary Table 10 reports the phase experts before distillation. The table below adds the same experts after distillation by Spatial-NPD (accuracy in %, before → after).

BenchmarkSamplesBaseShared foundationExploration expertApproach expert
BLINK Spatial40066.7569.2568.75 → 67.2569.75 → 69.50
Q-Spatial++10159.4172.2869.31 → 61.3971.29 → 68.32
Q-Spatial-ScanNet8264.6368.2971.95 → 70.7369.51 → 68.29
SpatialEval-VQA4,63543.8244.5146.04 → 43.3744.83 → 44.10
VSI-Bench5,13055.6255.6755.12 → 55.5355.69 → 55.79
OmniSpatial1,53343.1841.4240.64 → 42.9941.29 → 41.62

Across each branch, distillation reduces accuracy on four benchmarks while improving it on VSI-Bench and OmniSpatial. Eleven of the twelve score changes are within 3.0 points. The only notable exception is the Exploration expert on Q-Spatial++, a metric distance-estimation benchmark, where the number of correct answers drops from 70 to 62 out of 101. This observation is consistent with the trade-off you raised: distillation toward movement-oriented decisions may come at the expense of accuracy on explicit metric distance estimation.

We therefore maintain that Spatial-NPD improves the utilization of spatial evidence for navigation, while preserving most general spatial capabilities with only modest degradation. We will include the post-distillation results in Supplementary Table 10 and explicitly acknowledge this trade-off in the limitations.

W2. Meaning, evaluation, and examples of the ability types

Meaning. The specific meanings of the five abilities are as follows.

AbilityDefinitionExample annotation
VisibilityWhether the target or a relevant landmark can be visually confirmed"The TV cannot be confirmed because a closed door blocks the view."
TopologyWhether a visible transition connects regions"The staircase leads to a visible landing and doorway; connectivity beyond the doorway is unknown."
AffordanceWhether a visible local region appears usable for movement or interaction"The central floor strip appears traversable, but it is unclear whether there is enough clearance for the robot to pass through."
RelationObservable relative position among objects, the goal, and candidate regions"The refrigerator is near the doorway to the right of the staircase."
GeometryVisible shape, scale, and coarse distance cues"The target appears closer and large enough to consider stopping."

Each record states the ability, a claim, the visible evidence, its navigation relevance, and an evidence limit, and contains no action label (as structure in Appendix A.3):

{
  "ability": ["<ability a>", "<ability b>", ...],
  "spatial_reasoning": {
    "<ability a>": {
      "claim": "...",
      "visual_evidence": "...",
      "navigation_relevance": "...",
      "evidence_limit": "..."
    },
    ...
  }
}

Distribution. Of the 112,886 total records, 48,288 carry at least one spatial ability tag, and a record may carry several.

Spatial abilityShared foundationExplorationApproachTotal
Visibility11,145—9,78020,925
Geometry8,522—8,19116,713
Relation5,9767,4443,45016,870
Topology5,0607,142—12,202
Affordance4,1458,221—12,366
Tag occurrences34,84822,80721,42179,076
Ability-tagged records16,00022,5089,78048,288

The shared foundation covers all five abilities. Exploration uses the abilities that inform where to go next, and Approach uses those that inform whether the target is confirmed and close enough to stop.

Evaluation. We evaluate the abilities at three levels.

  • Annotation quality. An automatic audit of annotations checks output schema and evidence fields, action-label separation, and privileged-information boundaries; 99.93% pass and the rest are removed. A manual review of 500 random annotation-image pairs found 97.8% semantic correctness, 98.8% visual grounding, 98.4% ability-label appropriateness, and 98.2% conservative uncertainty handling, 98.3% on average.
  • Ability acquisition. For each ability, we compare the model's spatial response on a held-out observation with the held-out GPT-5.4 annotation and report the mean embedding cosine similarity (Figure 6(b), Supplementary Table 11). These annotations are not used for training.
  • Effect on navigation. Main Table 3 trains one model per ability and measures SR / SPL on HM3D-v0.2. Affordance and Topology help Exploration most (+5.2 and +2.8 SR), and Visibility and Geometry help Approach most (+3.1 and +3.0 SPL).

Examples. We added an image-annotation example for each ability to Supplementary Appendix A.3 (Figure 7). More examples are in the anonymous repository, which links the dataset.

Reviewer_VGmnOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how spatial reasoning can improve Object Goal Navigation (ObjectNav). Through systematic analysis, the authors identify the importance of goal-coupled spatial supervision and phase-specific spatial abilities. Based on these findings, they introduce Spatial-Nav-100K, a two-stage fine-tuning strategy, and Spatial-NPD for distilling spatial knowledge into navigation policies. Experiments on HM3D and MP3D demonstrate competitive performance with low policy training costs.

Critical strengths
  1. The paper addresses a meaningful research question by systematically investigating how spatial reasoning transfers to navigation. The analysis provides useful insights into goal-coupled supervision and phase-specific spatial abilities.

  2. The proposed framework achieves competitive navigation performance on HM3D and MP3D with relatively low policy training costs. Extensive ablation studies further support the effectiveness of the proposed components.

Critical weaknesses
  1. Limited evidence for the transferability of spatial reasoning. Table 1 shows that benchmark-oriented spatial specialists underperform their base models in navigation. However, it remains unclear whether this degradation reflects limited spatial reasoning transferability or incompatibility with navigation prompts and action interfaces. Further clarification is needed.

  2. Insufficient isolation of Spatial-NPD's contribution. Although Tables 6 and 7 demonstrate the effectiveness of Spatial-NPD, it remains unclear whether the gains primarily stem from self-distillation or the additional spatial supervision. Comparisons with standard self-distillation without spatial priors and continued SFT would help clarify its independent contribution.

Questions for authors

Could the authors provide more details about the construction of Spatial-Nav-100K, particularly the state sampling strategy, the distribution of spatial abilities and navigation phases, and the quality verification of GPT-generated annotations?

6 comments

Response to Reviewer VGmn

Conclusion

Thank you for the review. We conducted additional controlled experiments to disentangle spatial-reasoning transfer from prompt and action-interface mismatch, and to further validate the contribution of Spatial-NPD beyond continued SFT and prior-free self-distillation. The results demonstrate that interface adaptation only partially explains the performance gap, while the aligned spatial prior provides additional navigation gains beyond conventional training and distillation.

In addition, we provided more details of the proposed Spatial-Nav-100K dataset, including its construction protocol, spatial ability distribution, annotation pipeline, quality verification, and released the anonymized code repository and dataset.

W1. Limited evidence for the transferability of spatial reasoning

We acknowledge that Table 1 alone cannot disentangle limited spatial reasoning transferability from incompatibility with navigation prompts and action interfaces.

To distinguish these factors, we measured each model's action legality rate in WMNav on the HM3D-v0.2 benchmark and evaluated two training-free adaptations: Prompt-adapted, which clarifies action definitions, and Format-adapted, which constrains outputs to a structured action schema. Action legality rate measures the proportion of decision steps yielding a legal action. Results are reported as action legality rate / SR / SPL.

ModelOriginalPrompt-adaptedFormat-adapted
VST-SFT17.0% / 26.2 / 8.355.8% / 40.5 / 14.672.6% / 30.0 / 10.2
VST-RL45.6% / 49.3 / 16.1——
ViLASR15.4% / 30.1 / 9.6——
Video-R190.9% / 35.8 / 8.994.3% / 38.5 / 10.496.7% / 36.0 / 9.6
SenseNova-SI (Qwen2.5)100.0% / 55.3 / 19.2——
SenseNova-SI (Qwen3)100.0% / 59.8 / 22.5100.0% / 60.2 / 22.8100.0% / 59.9 / 22.6
Qwen3-VL-8B Base100.0% / 70.1 / 31.0100.0% / 70.8 / 31.4100.0% / 71.9 / 32.0

We ran the two adaptations on four models whose action legality rates range from 17.0% to 100%.

These adaptations improve action legality and, in some cases, navigation performance (e.g., VST-SFT improves from 26.2 to 40.5 SR with prompt adaptation). However, higher action legality does not necessarily yield better navigation (e.g., VST-SFT achieves only 30.0 SR with format adaptation despite 72.6% legality). More importantly, even with 100% action legality across all settings, SenseNova-SI consistently lags behind Qwen3 Base by 10.3–12.0 SR.

These results disentangle the two factors and demonstrate that interface incompatibility explains only part of the degradation: the fundamental limitation lies in the specialists' inability to effectively translate spatial cognition into navigation actions.

W2. Insufficient isolation of Spatial-NPD's contribution

We agree that the existing results do not fully isolate the contributions of self-distillation and spatial supervision. To address this, we added two controls: continued SFT and prior-free self-distillation, both using the same action examples and update budget as Spatial-NPD. We report them alongside the shuffled- and aligned-prior results already presented in Table 7 to further clarify the role of spatial supervision.

MethodSpatial priorDistillationHM3D-v0.2 SR / SPL
Train w/o Spatial-NPDNoneNo74.5 / 32.9
Continued SFTNoneNo75.1 / 33.4
Prior-free self-distillationNoneYes75.8 / 34.1
Spatial-NPD, shuffled priorShuffledYes75.0 / 32.6
Spatial-NPD, aligned priorAlignedYes77.4 / 35.4

Continued SFT and prior-free self-distillation yield gains of only +0.6/+0.5 and +1.3/+1.2 SR/SPL, respectively. In contrast, Spatial-NPD with aligned priors achieves +2.9/+2.5 over the baseline, further outperforming prior-free self-distillation by +1.6 SR / +1.3 SPL. These results demonstrate that self-distillation alone cannot account for the gains, and effective spatial supervision is essential to Spatial-NPD's improvement. We also incorporate these results into the paper and clarify the independent contribution of spatial supervision in §5.3.

Q. More details about the construction of Spatial-Nav-100K

Appendix A provides construction details for Spatial-Nav-100K. We further detail the state sampling strategy, distributions of spatial abilities and navigation phases, and quality verification of GPT-generated annotations. Code, configuration files, and audit scripts are in the anonymous repository, which links to Spatial-Nav-100K.

State sampling strategy. We first collect approximately 92K candidate navigation states (including metadata such as RGB and goal information) from Qwen3-VL-8B trajectories generated under the VLMNav framework on the training splits of HM3D-v0.1 and HM3D-v0.2, covering 5.25K ObjectNav episodes. We then construct the final sampled state set from this candidate pool by uniformly sampling across navigation statuses and target categories, subject to two constraints: at most six states are selected from each episode, and they are separated by at least five navigation steps. Each state can yield multiple supervision records (spatial reasoning, Spatial2Nav, Exploration, and Approach), totaling 112,886 records (107,442 training; 5,444 development). The shared foundation uses training records only (Supplementary Table 5); held-out states are excluded from training and model selection.

Distribution of spatial abilities and navigation phases. Supplementary Table 2 breaks down Spatial-Nav-100K by supervision type, navigation phase, and spatial ability. It contains 112,886 records: 40,000 in the shared foundation, 58,000 in Exploration, and 14,886 in Approach. The first table reports Navigation Decision and Spatial Reasoning counts; the second details ability tags. As records may carry multiple tags, 79,076 tag occurrences correspond to 48,288 unique ability-tagged records.

Supervision typeShared foundationExplorationApproachTotal
Navigation Decision24,00035,4925,10664,598
Spatial Reasoning16,00022,5089,78048,288
Total40,00058,00014,886112,886
Spatial abilityShared foundationExplorationApproachTotal
Visibility11,145—9,78020,925
Geometry8,522—8,19116,713
Relation5,9767,4443,45016,870
Topology5,0607,142—12,202
Affordance4,1458,221—12,366
Tag occurrences34,84822,80721,42179,076
Ability-tagged records16,00022,5089,78048,288

Annotation. Of the 112,886 records, 15,106 are Approach-phase decision records (10,000 in the shared foundation and 5,106 in the Approach stage). Simulator rules label them based on target distance and visibility, so no GPT annotation is needed. GPT-5.4 generates the remaining 97,780 from one RGB frame and public task information. Exploration decision proposals are retained only after simulator expert-rule verification. Each spatial record includes the ability, a claim, visual evidence, navigation relevance, and an evidence limit, but no action label:

{
  "ability": ["<ability a>", "<ability b>", ...],
  "spatial_reasoning": {
    "<ability a>": {
      "claim": "...",
      "visual_evidence": "...",
      "navigation_relevance": "...",
      "evidence_limit": "..."
    },
    ...
  }
}

The five abilities are defined as follows.

AbilityDefinitionExample annotation
VisibilityWhether the target or a relevant landmark can be visually confirmed"The TV cannot be confirmed because a closed door blocks the view."
TopologyWhether a visible transition connects regions"The staircase leads to a visible landing and doorway; connectivity beyond the doorway is unknown."
AffordanceWhether a visible local region appears usable for movement or interaction"The central floor strip appears traversable, but it is unclear whether there is enough clearance for the robot to pass through."
RelationObservable relative position among objects, the goal, and candidate regions"The refrigerator is near the doorway to the right of the staircase."
GeometryVisible shape, scale, and coarse distance cues"The target appears closer and large enough to consider stopping."

Supplementary Appendix A also includes these definitions, one image-annotation example per ability, the distribution table, and the four-criterion review breakdown.

Quality verification. Automatic audits check output schema/evidence fields, action-label separation, and privileged-information boundaries; 99.93% pass, and failures are removed. Manual review of 500 random annotation-image pairs found 97.8% semantic correctness, 98.8% visual grounding, 98.4% ability-label appropriateness, and 98.2% conservative uncertainty handling (98.3% average). These details are also added to the Appendix.

Thank you for the thoughtful response from the author. I have observed through the newly added experiments that even after the prompt and format adaptation, there is still a certain performance gap. Therefore, I believe that based on the current results, it is impossible to determine whether the remaining gap is due to the model's failure to apply spatial judgment to navigation decisions. Thus, I maintain the original division.

Reviewer_w8YqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how spatial supervision can improve object-goal navigation and proposes a method combining goal-coupled spatial annotations, a two-stage fine-tuning scheme that learns a shared spatial-navigation foundation before specializing separate exploration and stopping policies, and distillation from a teacher conditioned on spatial priors (Spatial-NPD).

Critical strengths

Two critical strengths are listed below.

  1. The comparison of supervision schemes keeps the visual states, annotator, and training settings fixed, and includes goal-oriented non-spatial supervision. This helps distinguish providing goal information from connecting that information to spatial evidence, making it one of the more convincing experiments in the paper.

  2. The appendix separately reports whether the agent reaches within one metre of the goal and whether it stops too early. Tuning exploration alone raises the arrival rate from 71.2% to 77.2%. These behavioral results provide more direct support for treating exploration and stopping differently than overall success rates alone.

Critical weaknesses

Three critical weaknesses are listed below.

  1. The Spatial-NPD comparison does not rule out benefits from ordinary distillation or additional training. In supplementary Table 6, shuffled priors reduce teacher action accuracy from 83.03% to 63.40%, so this is not a neutral control. Outperforming misleading priors does not establish an advantage over distillation without priors or further action-supervised training with the same budget.

  2. The gains from the training structure are still difficult to separate from the amount of training. Appendix C.1 states that the settings are shared, and supplementary Table 5 gives step counts only for the final pipeline, so total updates and sample exposure for the one-pass and joint variants are not listed. Since the shared-foundation design includes an additional training stage, these quantities need to be matched before attributing the gains specifically to shared learning and phase specialization.

  3. The abstract and conclusion state that spatial specialization tuned for benchmarks does not transfer to navigation, which extends beyond what the direct-replacement experiment establishes. Main Table 1 shows worse navigation results for the tested specialists, but Section 3.1 does not separate spatial limitations from action-format mismatch or reduced instruction following. It supports a failure of direct replacement in these settings, rather than a general explanation of why spatial specialization does not improve navigation.

Questions for authors

I list the questions below. These may influence my understanding of the method's details and experiment results.

  1. With the same action examples and update budget, how does Spatial-NPD compare with distillation from a teacher without spatial priors and with further action-supervised training alone?

  2. What are the total updates and sample exposure for each variant in supplementary Table 8? How does its one-pass result of 23.7 SPL differ in configuration from the phase-expert result of 28.1 SPL in main Table 6?

  3. How often do the specialists in Table 1 produce valid action outputs? Does the performance gap remain after adapting the action prompts or output format without further model training?

  4. Were the three abilities selected in Appendix C.2 chosen on development episodes separate from the final 1,000 HM3D-v0.2 evaluation episodes?

  5. How should the 226M input and 36M output tokens in Appendix B.3 be reconciled with the annotation counts of 100M/32M and audit counts of 54.5M/9M in supplementary Table 7?

3 comments

Response to Reviewer w8Yq

Conclusion

Thank you for the detailed review. We have conducted additional experiments and provided clarifications on the following aspects:

  • Spatial-NPD comparisons: Added continued SFT and prior-free self-distillation controls under matched training budgets.
  • Training budget: Reported optimizer updates and training samples processed for all variants and clarified the differences between SPL results.
  • Direct model replacement: Evaluated action legality rates and prompt/output-format adaptations, and narrowed our claims accordingly.
  • Ability selection: Clarified the phase-specific ability selection process and test-set isolation.
  • Token accounting: Distinguished planned token budgets from actual consumption and provided a detailed breakdown.

In addition, we have released the anonymized code repository and Spatial-Nav-100K dataset, including implementation code, data, and annotation scripts to facilitate reproducibility.

W1 / Q1. Spatial-NPD versus prior-free distillation and continued SFT

We agree with the reviewer. Since shuffled priors degrade teacher accuracy (83.03% to 63.40%), they serve as a misleading-prior control rather than a neutral one. To strictly isolate the efficacy of aligned spatial priors from ordinary distillation and extended training, we added two neutral controls: Continued SFT (further action-supervised training) and Prior-free Self-distillation (distillation without priors).

For a fair comparison, all extended variants start from the same base checkpoint (2,526 optimizer updates) and receive the exact same additional budget (417 optimizer updates on the identical 13,317 samples), totaling 2,943 optimizer updates across all methods:

MethodSpatial PriorDistillationOptimizer UpdatesHM3D-v0.2 SR (↑\uparrow) / SPL (↑\uparrow)
Train w/o Spatial-NPDNoneNo2,52674.5 / 32.9
Continued SFT (New)NoneNo2,94375.1 / 33.4
Prior-free Self-distillation (New)NoneYes2,94375.8 / 34.1
Spatial-NPD (Misleading Control)ShuffledYes2,94375.0 / 32.6
Spatial-NPDAlignedYes2,94377.4 / 35.4

These results support the contribution of aligned spatial priors beyond additional training and prior-free self-distillation. We also add these two neutral control experiments to the paper accordingly.

W2 / Q2. Training gains are confounded by different training budgets.

We clarify that training gains are not confounded by budget differences. Every variant in Supplementary Table 8 is strictly matched to the same total budget of 2,526 optimizer updates and 134,125 processed training samples (counting each sample once per epoch).

Specifically, our staged pipeline before Spatial-NPD (which achieves 74.5 SR / 32.9 SPL) distributes this budget across three stages as follows:

StageEpochsBatchOptimizer Updates
Shared foundation296834
Exploration1321,250
Approach132442
Total2,526

When an ablation variant omits a stage, its remaining stages are scaled up proportionally to maintain the exact 2,526 optimizer updates:

  • One-pass phase-specific: Trains the two phase policies directly, omitting both the shared-foundation stage and its data.
  • Joint Exploration/Approach: Trains a single policy for both phases (corresponding to "+ Decision-spatial SFT" in Table 6(a)).
  • Staged w/o spatial supervision: Replaces spatial targets with action supervision.
VariantShared-foundationPhase-specificSpatial SupervisionOptimizer UpdatesSR / SPL
Staged phase-specific (Ours)✓✓✓2,52674.5 / 32.9
One-pass phase-specific✗✓✓2,52668.8 / 23.7
Joint Exploration/Approach✗✗✓2,52663.7 / 26.5
Staged w/o spatial supervision✓✓✗2,52663.9 / 28.0

Under this strictly matched budget, our staged pipeline outperforms one-pass training by +5.7 SR / +9.2 SPL and the joint policy by +10.8 SR / +6.4 SPL. This confirms that the performance gains stem from our staged architecture and spatial supervision, rather than a larger training budget.

Difference between "One-pass" (23.7 SPL) and "+ Phase experts" (28.1 SPL): These two variants differ in whether they utilize the shared-foundation data. While both omit the separate shared-foundation training stage, "One-pass phase-specific" (23.7 SPL) completely discards the shared-foundation data. In contrast, "+ Phase experts" (28.1 SPL) retains the shared-foundation data and mixes it directly into phase-expert training. The 4.4-SPL difference therefore reflects whether the shared-foundation data is used, not whether a separate shared-foundation stage exists.

W3 / Q3. Overclaim about why spatial specialization fails in navigation

We agree that our original claim was too broad. To disentangle spatial reasoning transferability from compatibility with navigation prompts and action interfaces, we measured the action legality rate (the percentage of decision steps yielding valid, parseable actions) in WMNav on HM3D-v0.2. We further evaluated two training-free adaptations on four representative models spanning action legality rates from 17.0% to 100%: Prompt-adapted, which clarifies action definitions, and Format-adapted, which constrains outputs to a structured action schema.

ModelOriginalPrompt-adaptedFormat-adapted
VST-SFT17.0% / 26.2 / 8.355.8% / 40.5 / 14.672.6% / 30.0 / 10.2
VST-RL45.6% / 49.3 / 16.1——
ViLASR15.4% / 30.1 / 9.6——
Video-R190.9% / 35.8 / 8.994.3% / 38.5 / 10.496.7% / 36.0 / 9.6
SenseNova-SI (Qwen2.5)100.0% / 55.3 / 19.2——
SenseNova-SI (Qwen3)100.0% / 59.8 / 22.5100.0% / 60.2 / 22.8100.0% / 59.9 / 22.6
Qwen3-VL-8B Base100.0% / 70.1 / 31.0100.0% / 70.8 / 31.4100.0% / 71.9 / 32.0

Results are reported as action legality rate / SR / SPL. Interface adaptation partially recovers navigation performance for models with low legality rates. For example, prompt adaptation improves VST-SFT's SR from 26.2 to 40.5, although it remains below its base model (58.2 SR). However, higher legality does not guarantee better navigation: format adaptation raises VST-SFT's legality to 72.6% but yields only 30.0 SR. More importantly, despite 100% legality across all settings, SenseNova-SI (Qwen3) consistently underperforms Qwen3 Base by 10.3–12.0 SR.

These results separate interface incompatibility from the persistent navigation performance gap. Interface mismatch explains part of the degradation, but the fundamental limitation lies in spatial specialists' inability to effectively translate spatial cognition into navigation actions. We also add this analysis to paper and revise the corresponding claims.

Q4. Were the abilities selection separate from HM3D-v0.2 evaluation episodes?

Yes, the final 1,000 HM3D-v0.2 evaluation episodes were strictly excluded from ability selection, phase assignment, hyperparameter tuning, checkpoint selection, and prompt/format design. Moreover, HM3D-v0.1 and MP3D were completely excluded from development and served solely for external generalization evaluation.

The five ability families and their phase assignments were determined a priori using only the training and development splits, based on the functional requirements of each phase (Figure 1): Exploration uses Relation, Topology, and Affordance to guide movement, while Approach uses Visibility, Geometry, and Relation to verify target visibility and proximity before stopping. All assignments were frozen prior to final test evaluation.

Table 3 is a post hoc analysis conducted after the assignments were frozen, rather than the basis for ability selection or model development. Notably, although Geometry slightly improves Exploration in Table 3 (+0.8 SPL), it was not included in the pre-frozen Exploration set, which is consistent with this order.

We acknowledge that the phrasing "retained according to the Table 3 gains" in Appendix C.2 inadvertently implies the reverse order. We also correct this wording and explicitly clarify the development timeline and test-set exclusion in Appendix C.2 and Section 3.2.2.

Q5. Confusion about inconsistent token counts.

We apologize for the confusion caused by our ambiguous phrasing. The two sets of numbers differ in scope: the figures in Supplementary Table 7 (100M/32M and 54.5M/9M) represent our initial estimated budgets for annotation and auditing, whereas the 226M/36M in Appendix B.3 represents the total realized usage across the entire pipeline.

The exact breakdown comparing planned budgets against realized token usage is detailed below:

WorkflowPlanned Input / OutputRealized Input / Output
Annotation, including retries100.0M / 32.0M162.55M / 28.50M
Automatic audit54.5M / 9.0M63.57M / 7.77M
Total154.5M / 41.0M226.12M / 36.27M

We have revised Appendix B.3 and Supplementary Table 7 to explicitly distinguish planned budgets from realized usage and include this complete breakdown.

We thank the reviewer once again for their valuable feedback. We have addressed all the points raised and incorporated the suggestions to further strengthen our paper. With only about a day remaining in the discussion period, we would like to kindly ask whether the reviewer has any remaining concerns that we could address before the deadline.

Sincerely,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.