acceptodds
Under review as a conference paper at ICLR 2027

Semantic Coverage and Temporal Selection for Visual Token Pruning in Vision-Language Navigation

Abstract

Vision-language navigation (VLN) enables embodied agents to follow natural-language instructions in real-world environments. However, dense visual tokens incur substantial computation and inference latency, hindering deployment on resource-constrained robots. Existing VLN token-pruning methods neither explicitly suppress redundancy across historical frames nor exploit instruction semantics during selection. We propose a training-free spatio-temporal visual-token compressor that captures intra-frame semantic diversity and inter-frame novelty for efficient long-horizon inference. For the current frame, low-rank feature grouping identifies complementary semantic groups, while adaptive within-group selection uses feature-space non-maximum suppression to remove redundant tokens without sacrificing semantic coverage. For historical frames, we adopt the same strategy as current frame to construct an expanded candidate pool. Then, we assess each token using current-view relevance, instruction alignment, and visual saliency. Moreover, to reduce temporal redundancy, semi-relaxed temporal alignment uses optimal transport to measure inter-frame novelty against preceding-frame group prototypes. Task-aware token selection uses the combined scores and maximal marginal relevance to iteratively select the final token set. Experiments on standard VLN benchmarks show that our method consistently outperforms competing pruning baselines across compression ratios, achieving a better navigation performance and efficiency trade-off. Real-world quadruped robot experiments further validate its effectiveness for low-latency embodied navigation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Visual Token Pruning, Vision-Language Navigation, Semi-relaxed Optimal Transport

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_p3cVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes training-free visual-token compression for navigation. It combines low-rank feature grouping and within-group selection with instruction relevance, temporal novelty from semi-relaxed optimal transport, and diversity-aware historical selection. Evaluation covers R2R-CE, RxR-CE, and additional deployment examples.

Critical strengths
  • The reported SPL improvements over A-MMR are consistent across both benchmarks and all three pruning ratios. The stronger gains under aggressive compression suggest that the method addresses a meaningful limitation of uniform or saliency-dominated selection.

  • The component ablations are reasonably comprehensive and identify grouping and instruction relevance as useful components. The additional NaVILA evaluation also provides some evidence of transfer beyond the main StreamVLN backbone.

Critical weaknesses
  • The main comparative evidence is not fully controlled. Baseline results are imported from A-MMR, and equivalence of checkpoints, history settings, and decoding configurations is not demonstrated. The appendix acknowledges this limitation, but that qualification does not resolve the uncertainty behind the main improvement claims. A-MMR is also absent from the efficiency comparison despite being the most relevant baseline.

  • Important implementation and measurement details remain unresolved. The threshold in Equation 5 can exceed 1, in which case the cosine-based suppression accepts every candidate. Its actual range and any clipping therefore matter. Table 3 additionally reports 4.43 FPS alongside 248.26 ms latency without explaining the timing definitions; the other pruning rows approximately follow the reciprocal relationship.

Questions for authors
  • Is the threshold in Equation 5 clipped or normalized? What values occur in practice, and how often does within-group suppression remove tokens?

  • Can you reproduce A-MMR under matched conditions, include all selection preprocessing in timing, and reconcile the FPS and latency measurements?

5 comments
AuthorsOpenReview ↗

Response to Reviewer p3cV (1/2)

We sincerely thank the reviewer for the careful evaluation and constructive comments, particularly regarding comparisons, threshold, and efficiency. We have carefully revisited these issues, conducted additional experiments, and corrected the relevant errors in the manuscript. Below, we address each concern in detail.


W1: The main comparative evidence is not fully controlled.

A1: Thank you for raising this point. We clarify the evaluation setup and provide additional results to address the concern.

  • A-MMR reproduction. To ensure a consistent and fair comparison, we made every effort to reproduce A-MMR based on its published paper, as shown in Tab. R1. Our reproduced results closely match the originally reported performance, and even surpass in some metrics, further supporting the reliability and fairness of our comparison. Since the official implementation is unavailable, we carefully aligned the model checkpoint (StreamVLN_Video_qwen_1_5_r2r_rxr_envdrop_scalevln_v1_3), evaluator (Habitat 0.2.4), history configuration (8 uniformly sampled frames), decoding settings, and all other publicly specified configurations. Only a few undisclosed inference settings could not be exactly replicated.

    Table R1: Navigation performance on the R2R-CE validation-unseen split. The symbols ′ and ″ denote results reported in original paper and those obtained from our reproduction, respectively. To ensure a fair comparison, we maintain consistency with the original work in all publicly specified settings, including model checkpoints, evaluators, and decoding configurations, while some undisclosed implementation details cannot be exactly replicated. Bold indicates the best results.

    PruningMethodSR ↑SPL ↑OS ↑NE ↓
    0% (729)Unpruned′55.7449.6663.605.06
    Unpruned″54.8749.3362.155.21
    70% (218)A-MMR′51.5047.2858.025.56
    A-MMR″52.8549.4458.455.28
    Ours55.2550.5961.075.04
    80% (146)A-MMR′52.5846.4460.095.38
    A-MMR″53.9948.5561.504.92
    Ours57.3750.4664.824.71
    90% (72)A-MMR′47.6336.3668.465.69
    A-MMR″46.5537.0367.165.82
    Ours51.7739.1071.725.23
  • Efficiency. We have also incorporated the reproduced A-MMR into the efficiency comparison and extended the evaluation to 70%, 80%, and 90% pruning, as shown in Tab. R2. Our method consistently achieves higher throughput and lower inference latency than the reproduced A-MMR, while maintaining comparable computational costs.

    Table R2: Efficiency analysis of inference throughput (FPS), computational cost (TFLOPs), and CUDA inference latency on the StreamVLN backbone using an NVIDIA A100 GPU. FPS measures inference throughput, TFLOPs reflects computational cost per inference, and Latency measures model inference time in milliseconds. The best and second-best results among pruning methods at each pruning ratio are highlighted in bold and italics, respectively.

    PruningMethodFPS ↑TFLOPs ↓Latency (ms) ↓
    0% (729)Unpruned3.143.76318.34
    70% (218)SparseVLM3.199.69313.16
    DivPrune5.094.08196.31
    VisPruner5.154.06194.28
    A-MMR4.644.06215.38
    Ours5.054.08197.92
    80% (146)SparseVLM3.968.67252.32
    DivPrune5.333.06187.54
    VisPruner5.073.04197.13
    A-MMR3.833.04260.89
    Ours5.103.06196.29
    90% (72)SparseVLM4.037.62248.26
    DivPrune4.682.01213.64
    VisPruner4.042.00247.57
    A-MMR3.651.99274.31
    Ours4.432.01225.62

Together with the reproduced navigation results, these experiments provide stronger evidence of the favorable navigation–efficiency trade-off achieved by our method.


AuthorsOpenReview ↗

Response to Reviewer p3cV (2/2)


W2: Important implementation and measurement details remain unresolved.

A2: We appreciate the reviewer for identifying these issues. After carefully re-examining the formulation and experimental records, we have made the following corrections.

  • Eq. (5). The normalization factor in Eq. (5) is 1/N1/N. The correct formulation is: θB=BN2(N−1)∑i≠jcos⁡(xi,xj),\theta_B=\frac{B}{N^2(N-1)}\sum_{i\neq j}\cos(x_i,x_j), where B/NB/N represents the token retention ratio. Since B<NB<N, the resulting threshold remains below 1, avoiding the issue raised by the reviewer. As the pruning ratio increases, the threshold becomes stricter, while proportional budget allocation ensures group coverage and the required output budget.

  • Tab. 3. We sincerely apologize for the error in the reported latency value in Tab. 3. The correct inference latency of our method at 90% pruning is 225.62 ms, corresponding to 4.43 FPS. We have further clarified the definitions and protocols of all efficiency metrics and extended the comparison to 70%, 80%, and 90% pruning, as presented in Tab. R2.

We apologize for these reporting errors and have corrected the corresponding descriptions in the revised manuscript.


Q1: Is the threshold in Equation 5 clipped or normalized? What values occur in practice, and how often does within-group suppression remove tokens?

A1: The threshold in Eq. 5 is normalized by the token count NN. As explained in our response to W2, the corrected formulation ensures that θB<1\theta_B<1. To further clarify the practical behavior of within-group suppression, we analyzed 10 navigation trajectories under different pruning ratios. For each trajectory, we measured the average threshold θB\theta_B and token suppression rate across all navigation steps. The results are as follows:

Table R3. Average adaptive threshold θB\theta_B and within-group token suppression rate across 10 navigation trajectories under different pruning ratios. Suppression rates are measured before budget allocation.

PruningStageSuppression Rate (%)Mean θB\theta_B
90% (72)Current frame96.450.0259
Historical frames93.350.0532
80% (146)Current frame93.430.0527
Historical frames86.040.1065
70% (218)Current frame90.040.0785
Historical frames78.360.1581

The results show that the average threshold ranges from 0.03 to 0.16, while within-group suppression removes 78.36%–96.45% of input tokens before budget allocation. These results demonstrate that the suppression mechanism effectively removes redundant tokens across different pruning ratios.

Q2: Can you reproduce A-MMR under matched conditions, include all selection preprocessing in timing, and reconcile the FPS and latency measurements?

A2: We have reproduced A-MMR under matched experimental conditions and provided an extended efficiency comparison. Please refer to our response to W1 for the detailed results.


Dear Reviewer p3cV,

Thank you again for your careful review and constructive suggestions. We sincerely appreciate the time and effort you have devoted to evaluating our work.

We have provided point-by-point responses, additional experiments, and clarifications in our rebuttal. As the discussion period draws to a close, we would be grateful if you could let us know whether these responses have addressed your concerns and which issues, if any, remain unresolved. We welcome further discussion and would be happy to provide timely clarifications before the deadline.

Thank you for your careful consideration and the effort you have invested in helping us improve our work.

Best regards,

The Authors

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

The paper reports consistent navigation improvements and provides useful component ablations. However, the practical advantage over A-MMR remains insufficiently established without a matched efficiency comparison that clearly accounts for compression overhead. The implementation and practical behavior of the within-group selection rule also require clarification. These gaps limit confidence in the claimed performance–efficiency advantage, and I therefore lean toward rejection in the current form.

Critical strengths

Consistent reported improvements with supporting ablations. Table 1 reports higher SPL than A-MMR on both R2R-CE and RxR-CE at all three token budgets, averaging 3.30 percentage points of improvement. Table 2 further provides component ablations supporting the contributions of grouping, instruction relevance, and temporal novelty.

Critical weaknesses
  1. The practical advantage over A-MMR is not established. Table 3 omits this closest baseline and does not specify the token budget or whether the full compression overhead is included. Without a matched comparison, it remains unclear whether the additional computational complexity yields a favorable performance–latency trade-off.

  2. The behavior of a core pruning component is unclear. The threshold in Eq. (5) can exceed one, which would disable within-group redundancy suppression. The paper does not report whether this occurs in practice or specify the retained SVD rank, leaving important questions about the component’s behavior and reproducibility unresolved.

  3. The practical benefit on the second backbone is less clear. Although the method outperforms the pruning baselines on NaVILA, Table 5 reports 30.40% SR at 70% pruning, compared with 54.00% for the reported unpruned reference. Without corresponding efficiency measurements, the value of this substantial performance trade-off remains unclear.

Questions for authors
  1. How does the method compare with A-MMR in navigation performance and end-to-end latency under matched token budgets and implementation settings, including all compression costs? Please specify the budget and timing scope used in Table 3.

  2. Is Eq. (5) implemented exactly as written, and how frequently does its threshold reach or exceed one? Please report the observed suppression rate and specify the retained SVD rank.

  3. Table 1 uses baseline results from prior work. Can the reported gains be confirmed with A-MMR and the unpruned model evaluated under the same checkpoint, evaluator, and decoding settings as the proposed method?

6 comments
AuthorsOpenReview ↗

Response to Reviewer uCmY (1/3)

We sincerely thank the reviewer for the insightful comments and constructive suggestions. In response to the feedback from all reviewers, we have carefully revised the manuscript, with all changes highlighted in blue. We address each concern point by point below.


W1: The practical advantage over A-MMR is not established.

A1: Thank you for your valuable comment regarding the practical efficiency advantage over A-MMR. Our method achieves a favorable trade-off between navigation performance and inference efficiency compared with A-MMR. Since the official A-MMR implementation has not been released, we initially excluded it from the efficiency comparison to avoid potential inconsistencies. To provide a more rigorous comparison, we made every effort to reproduce A-MMR based on its published paper and extended the efficiency evaluation to all three pruning ratios (70%, 80%, and 90%), as shown in Tab. R1:

Table R1: Efficiency analysis of inference throughput (FPS), computational cost (TFLOPs), and CUDA inference latency on the StreamVLN backbone using an NVIDIA A100 GPU. FPS measures inference throughput, TFLOPs reflects computational cost per inference, and Latency measures model inference time in milliseconds. The best and second-best results among pruning methods at each pruning ratio are highlighted in bold and italics, respectively.

PruningMethodFPS ↑TFLOPs ↓Latency (ms) ↓
0% (729)Unpruned3.143.76318.34
70% (218)SparseVLM3.199.69313.16
DivPrune5.094.08196.31
VisPruner5.154.06194.28
A-MMR4.644.06215.38
Ours5.054.08197.92
80% (146)SparseVLM3.968.67252.32
DivPrune5.333.06187.54
VisPruner5.073.04197.13
A-MMR3.833.04260.89
Ours5.103.06196.29
90% (72)SparseVLM4.037.62248.26
DivPrune4.682.01213.64
VisPruner4.042.00247.57
A-MMR3.651.99274.31
Ours4.432.01225.62

Our method consistently achieves higher throughput and lower inference latency than A-MMR across all three pruning ratios, with comparable computational costs. Specifically, our method reduces inference latency by 8.1%, 24.8%, and 17.7% at 70%, 80%, and 90% pruning, respectively. These results further confirm that our method achieves a favorable trade-off between navigation performance and inference efficiency.


W2: The behavior of a core pruning component is unclear.

A2: Thank you for pointing out these important implementation details. We have carefully examined both issues and provide the following clarifications.

  • Adaptive threshold. We apologize for inadvertently omitting the normalization factor 1/N1/N in Eq. (5). The correct formulation is: θB=BN2(N−1)∑i≠jcos⁡(xi,xj),\theta_B=\frac{B}{N^2(N-1)}\sum_{i\neq j}\cos(x_i,x_j), where B/NB/N represents the token retention ratio. Since B<NB<N, the threshold remains below 1. As the pruning ratio increases, the threshold becomes stricter, while proportional budget allocation ensures group coverage and the required output budget. To further verify its practical effectiveness, we analyzed 10 navigation trajectories and measured the average threshold θB\theta_B and within-group suppression rate across all navigation steps under different pruning ratios, as shown in Tab. R2:

    Table R2. Average adaptive threshold θB\theta_B and within-group token suppression rate across 10 navigation trajectories under different token budgets. Suppression rates are measured before budget allocation.

    PruningStageSuppression Rate (%)Mean θB\theta_B
    90% (72)Current frame96.450.0259
    Historical frames93.350.0532
    80% (146)Current frame93.430.0527
    Historical frames86.040.1065
    70% (218)Current frame90.040.0785
    Historical frames78.360.1581

    The results show that the average threshold ranges from 0.026 to 0.158, while within-group suppression removes 78.36%-96.45% of input tokens before budget allocation. This confirms that the suppression mechanism actively removes redundant tokens across different pruning ratios. Meanwhile, proportional budget allocation ensure group coverage and the required output budget.

  • SVD rank. The retained rank is determined by the final token budget KK, with R0=⌊K/4⌋R_0=\lfloor K/4\rfloor for current frames and R0=⌊K/2⌋R_0=\lfloor K/2\rfloor for historical frames. For example, at 90% pruning (K=72K=72), the corresponding ranks are 18 and 36, respectively. We have clarified these settings in Appendix B.1.2 to improve reproducibility.


AuthorsOpenReview ↗

Response to Reviewer uCmY (2/3)


W3: The practical benefit on the second backbone is less clear.

A3: Thank you for your generous suggestion. Our results demonstrates that the proposed method can be effectively applied to different navigation backbones. Despite the greater sensitivity of NaVILA to token pruning, our method consistently preserves better navigation performance than baselines and provides measurable efficiency improvements. To further evaluate the efficiency of our method, we conducted additional experiments on NaVILA, as shown in Tab. R3:

Table R3. Inference efficiency comparison on the NaVILA backbone under different token budgets. We report inference throughput (FPS), computational cost (TFLOPs), and latency (ms). The best and second-best results among pruning methods within each budget are highlighted in bold and italics, respectively.

PruningMethodFPS ↑TFLOPs ↓Latency (ms) ↓
0%Unpruned1.6033.20625.46
70% (218)DivPrune0.9633.461047.17
VisPruner0.9933.461014.25
Ours0.9733.951035.21
80% (146)DivPrune1.2023.44834.60
VisPruner1.2323.44814.45
Ours1.2123.91823.63
90% (72)DivPrune1.6613.51602.78
VisPruner1.6413.51611.31
Ours1.7313.96578.16

At the tightest budget of 72 tokens, our method achieves the lowest latency of 578.16 ms, reducing inference latency by 7.6% and computational cost by 58.0% compared with unpruned NaVILA. It also achieves the highest throughput among the pruning methods. At larger budgets, our method maintains competitive efficiency.

Regarding for Performance degradation. NaVILA is not directly compatible with our token-pruning framework due to fundamental differences in their visual processing pipelines. NaVILA uses a dedicated downsampling module to compress visual features before passing them to the language model. To enable token pruning, we modified its original downsampling strategy (i.e., restoring the original 729 tokens instead of 196) and applied pruning before the language model, consistent with our implementation in StreamVLN. However, this modification inevitably changes the visual representations that the pretrained model was originally designed to process. As a result, the alignment between visual features and the navigation policy may be disrupted, leading to substantial performance degradation without retraining. Nevertheless, our method consistently outperforms other pruning baselines under the same adapted architecture, demonstrating its effectiveness and adaptability across different navigation backbones.


AuthorsOpenReview ↗

Response to Reviewer uCmY (3/3)


Q1: How does the method compare with A-MMR in navigation performance and end-to-end latency under matched token budgets and implementation settings, including all compression costs? Please specify the budget and timing scope used in Table 3.

A1: We reproduced A-MMR under matched settings and compared efficiency across all three pruning ratios. Please refer to W1 for details.

Q2: Is Eq. (5) implemented exactly as written, and how frequently does its threshold reach or exceed one? Please report the observed suppression rate and specify the retained SVD rank.

A2: We corrected Eq. (5), verified that the threshold remains below 1, and reported suppression rates and SVD ranks. Please refer to W2 for details.

Q3: Table 1 uses baseline results from prior work. Can the reported gains be confirmed with A-MMR and the unpruned model evaluated under the same checkpoint, evaluator, and decoding settings as the proposed method?

A3: We appreciate the reviewer's emphasis on fair comparisons. To address this concern, we made every effort to reproduce both A-MMR and the unpruned StreamVLN under the same model checkpoint (StreamVLN_Video_qwen_1_5_r2r_rxr_envdrop_scalevln_v1_3), evaluator (Habitat 0.2.4), history configuration (8 uniformly sampled frames), decoding settings, and all other publicly specified configurations. Only a few undisclosed inference settings could not be exactly replicated. The reproduced results are compared with the published results below.

Table R4. Navigation performance on the R2R-CE validation-unseen split under matched evaluation settings. The symbols ′ and ″ denote results reported in the original A-MMR paper and those obtained from our reproduction, respectively. Bold indicates the best results.

PruningMethodSR ↑SPL ↑OS ↑NE ↓
0% (729)Unpruned′55.7449.6663.605.06
Unpruned″54.8749.3362.155.21
70% (218)A-MMR′51.5047.2858.025.56
A-MMR″52.8549.4458.455.28
Ours55.2550.5961.075.04
80% (146)A-MMR′52.5846.4460.095.38
A-MMR″53.9948.5561.504.92
Ours57.3750.4664.824.71
90% (72)A-MMR′47.6336.3668.465.69
A-MMR″46.5537.0367.165.82
Ours51.7739.1071.725.23

The reproduced unpruned results closely match the published values. Together with the A-MMR reproduction, these controlled evaluations provide further evidence supporting the reliability of our comparisons and the navigation improvements achieved by our method.


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

This paper proposes a training-free visual-token pruning framework for VLN, combining intra-frame semantic coverage with instruction-aware historical token selection and prototype-based semi-relaxed OT for temporal novelty estimation. Experiments demonstrate promising navigation improvements over existing pruning methods across different compression ratios.

Critical strengths

1.The paper addresses a meaningful efficiency problem in VLN, with a clear motivation for preserving instruction-relevant information while reducing visual redundancy.

  1. The method achieves promising navigation performance on R2R-CE and RxR-CE across different pruning ratios, supported by useful ablation studies.

3.Real-world demonstrations on a Unitree Go2 robot further illustrate its potential applicability.

Critical weaknesses
  1. Incomplete navigation-efficiency comparison. Although Table 1 demonstrates promising navigation improvements, Table 3 omits A-MMR, one of the most relevant VLN pruning baselines. This makes it difficult to assess the claimed navigation-efficiency advantage over the closest prior work. A direct efficiency comparison with A-MMR under identical settings, including token-selection overhead, would strengthen this claim.

  2. Insufficient justification of key methodological designs. Table 2 compares semi-relaxed OT with balanced OT and removing temporal novelty, but does not establish its advantage over simpler prototype-matching strategies, such as nearest-prototype cosine distance. Such a comparison would help justify its additional complexity. Moreover, the adaptive threshold in Eq. (5) may exceed 1, potentially rendering within-group suppression ineffective. Clarification of its actual behavior is needed.

  3. Limited quantitative real-world evaluation. Section 4.4 and Appendix B.2.3 mainly present qualitative trajectories without reporting success rates or end-to-end latency. Moreover, policy inference runs on a remote RTX 4090 workstation. While these demonstrations show feasibility, they provide limited evidence for low-latency robotic deployment. Quantitative evaluation would strengthen the practical claims.

Overall, the paper addresses a well-motivated problem and demonstrates promising navigation improvements. However, the incomplete comparison with A-MMR weakens the claimed navigation-efficiency advantage, while several key methodological choices require further justification. The limited quantitative real-world evaluation also leaves the deployment benefits insufficiently established.

Questions for authors

1.The numerical value of the retained SVD rank (R_0) is not reported in Appendix B.1.2. Could the authors provide this hyperparameter for reproducibility?

2.Equation (10) uses similarity over instruction embeddings, whereas Appendix C describes a single representation of the complete instruction. Could the authors clarify the actual instruction representation used?

3.To what extent is the proposed framework specifically designed for VLN, rather than general instruction-conditioned video or VLA token pruning? Clarifying which components are VLN-specific and which are broadly applicable would help establish the scope of the contribution.

4 comments
AuthorsOpenReview ↗

Response to Reviewer jNBX (1/3)

We sincerely thank the reviewer for recognizing the motivation, clarity, promising results, and practical potential of our work. We greatly appreciate the constructive comments, which have helped us further strengthen the experimental evaluation and clarify several methodological details. We have carefully revised the manuscript, with all changes highlighted in blue. We address each concern below.


W1: Incomplete navigation-efficiency comparison.

A1: We fully agree that a direct comparison with A-MMR is important for evaluating the navigation–efficiency trade-off. Our additional experiments provide stronger evidence that our method achieves a more favorable balance between navigation performance and inference efficiency.

Since the official A-MMR implementation has not been released, we initially excluded it to avoid potentially unfair comparisons. To address this concern, we made every effort to reproduce A-MMR based on its published paper. As shown in Tab. R1, our reproduced navigation performance closely matches the published results, supporting the reliability of our comparison. We further included the reproduced A-MMR in the efficiency evaluation and extended the comparison to all three pruning ratios (70%, 80%, and 90%), as presented in Tab. R2. The results show that our method achieves lower inference latency and higher throughput than A-MMR, with comparable computational costs. Together, these experiments further substantiate the favorable navigation–efficiency trade-off of our method.

Table R1: Navigation performance on the R2R-CE validation-unseen split. The symbols ′ and ″ denote results reported in A-MMR and those obtained from our reproduction, respectively. To ensure a fair comparison, we maintain consistency with the original work in all publicly specified settings, including model checkpoints, evaluators, and decoding configurations, while some undisclosed parameters and implementation details cannot be exactly replicated. Bold indicates the best results.

PruningMethodSR ↑SPL ↑OS ↑NE ↓
0% (729)Unpruned′55.7449.6663.605.06
Unpruned″54.8749.3362.155.21
70% (218)A-MMR′51.5047.2858.025.56
A-MMR″52.8549.4458.455.28
Ours55.2550.5961.075.04
80% (146)A-MMR′52.5846.4460.095.38
A-MMR″53.9948.5561.504.92
Ours57.3750.4664.824.71
90% (72)A-MMR′47.6336.3668.465.69
A-MMR″46.5537.0367.165.82
Ours51.7739.1071.725.23

Table R2: Efficiency analysis of inference throughput (FPS), computational cost (TFLOPs), and CUDA inference latency on the StreamVLN backbone using an NVIDIA A100 GPU. The best and second-best results among pruning methods at each pruning ratio are highlighted in bold and italics, respectively.

PruningMethodFPS ↑TFLOPs ↓Latency (ms) ↓
0% (729)Unpruned3.143.76318.34
70% (218)SparseVLM3.199.69313.16
DivPrune5.094.08196.31
VisPruner5.154.06194.28
A-MMR4.644.06215.38
Ours5.054.08197.92
80% (146)SparseVLM3.968.67252.32
DivPrune5.333.06187.54
VisPruner5.073.04197.13
A-MMR3.833.04260.89
Ours5.103.06196.29
90% (72)SparseVLM4.037.62248.26
DivPrune4.682.01213.64
VisPruner4.042.00247.57
A-MMR3.651.99274.31
Ours4.432.01225.62

AuthorsOpenReview ↗

Response to Reviewer jNBX (2/3)


W2: Insufficient justification of key methodological designs.

A2: We thank the reviewer for highlighting these important concerns.

  • We introduce three lightweight matching strategies based on dot product, cosine similarity, and Euclidean distance. As shown in Tab. R3, our semi-relaxed OT consistently delivers superior navigation performance while maintaining comparable inference efficiency, further validating its effectiveness.

Table R3. Comparison of navigation performance and inference efficiency between our full method and three simpler prototype-matching strategies on the R2R-CE validation-unseen split. Dot., Cos., and Euc. denote nearest-prototype matching based on dot product, cosine similarity, and squared Euclidean distance, respectively. Latency measures model inference time. The best and second-best results within each pruning ratio are highlighted in bold and italics, respectively.

PruningSetupsSR ↑SPL ↑OS ↑NE ↓FPS ↑Latency (ms) ↓
70% (218)Dot.51.7147.6058.245.394.118242.819
Cos.53.7249.2759.825.093.945253.509
Euc.54.2149.8760.694.995.16193.783
Ours55.2550.5961.075.045.052197.92
80% (146)Dot.56.5049.6864.224.813.998250.119
Cos.56.6650.1564.494.714.595217.646
Euc.56.9950.3364.444.713.619276.307
Ours57.3750.4664.824.715.095196.29
90% (72)Dot.48.4036.6570.755.524.726211.591
Cos.49.3238.1268.905.554.808208.007
Euc.49.4337.5168.955.414.651215.019
Ours51.7739.1071.725.234.432225.62
  • We apologize for inadvertently omitting the normalization factor 1/N1/N in Eq. (5). The correct formulation is: θB=BN2(N−1)∑i≠jcos⁡(xi,xj),\theta_B=\frac{B}{N^2(N-1)}\sum_{i\neq j}\cos(x_i,x_j), where B/NB/N represents the token retention ratio. Since B<NB<N, the resulting threshold remains below 1, avoiding the issue raised by the reviewer. As pruning becomes more aggressive, B/NB/N decreases, enforcing stricter within-group selection. Meanwhile, proportional budget allocation ensure group coverage and the required token budget.

W3: Limited quantitative real-world evaluation.

A3: We appreciate the reviewer's suggestion and have conducted additional quantitative real-world experiments.

  • We evaluated 4 navigation trajectories, each with 3 different instructions, measuring navigation success rates and inference latency. As shown in Tab. R4, our method achieves the same success rate as the unpruned model (83.33%), while reducing inference latency from 285.29 ms to 272.72 ms. It also outperforms reproduced A-MMR in both navigation success rate and model inference efficiency.

Table R4. Quantitative real-world evaluation on 4 navigation trajectories under 90% pruning ratio (i.e., 72 tokens), each with 3 different instructions (12 trials per method). All methods are evaluated under the same deployment configuration. We report inference latency (ms) and navigation success rate (%), where inference latency measures model inference time and success rate denotes the percentage of successfully completed navigation trials.

MethodsInference Latency (ms) ↓Success Rate (%) ↑
Unpruned285.2983.33
A-MMR278.8975.00
Ours272.7283.33
  • Onboard inference remains challenging for large-scale VLN models on resource-constrained robots. Our deployment follows common strategy as follow StreamVLN[R1], DecoVLN[R2] and JanusVLN[R3], using a remote RTX 4090 workstation for all compared methods. Notably, all methods additionally share the same pre-inference and action execution pipeline, with an average latency of 24.44 ms from instruction issuance to model input and 1,072.49 ms from model output to action completion.

References

[R1] Wei M, et al. "Streamvln: Streaming vision-and-language navigation via slowfast context modeling." ICRA 2026.

[R2] Xin Z, et al. "Decovln: Decoupling observation, reasoning, and correction for vision-and-language navigation." CVPR 2026.

[R3] Zeng S, et al. "Janusvln: Decoupling semantics and spatiality with dual implicit memory for vision-language navigation." ICLR 2026.


AuthorsOpenReview ↗

Response to Reviewer jNBX (3/3)


Q1: The numerical value of the retained SVD rank (R_0) is not reported in Appendix B.1.2. Could the authors provide this hyperparameter for reproducibility?

A1: We thank the reviewer for pointing out this omission. The retained SVD rank is determined by the final per-frame token budget KK, with R0=⌊K/4⌋R_0=\lfloor K/4\rfloor for the current frame and R0=⌊K/2⌋R_0=\lfloor K/2\rfloor for historical frames, where ⌊⋅⌋\lfloor\cdot\rfloor denotes rounding down. For example, at 90% pruning (K=72K=72), the corresponding ranks are 18 and 36, respectively. We have clarified this setting in Appendix B.1.2 to ensure reproducibility.

Q2: Equation (10) uses similarity over instruction embeddings, whereas Appendix C describes a single representation of the complete instruction. Could the authors clarify the actual instruction representation used?

A2: We apologize for the ambiguous wording in Appendix C. Our method encodes the complete instruction as a sequence of token embeddings, as defined in Eq. (10), rather than separately processing individual sub-instructions. The term “single” was intended to emphasize that the instruction is treated as a whole, rather than represented by a single pooled embedding. We have revised the corresponding descriptions in Line 319 and Appendix C to eliminate this ambiguity.

Q3: To what extent is the proposed framework specifically designed for VLN, rather than general instruction-conditioned video or VLA token pruning? Clarifying which components are VLN-specific and which are broadly applicable would help establish the scope of the contribution.

A3: We thank the reviewer for this insightful question. Our framework differs from general video and VLA token pruning in the following aspects:

  • Compared with general video token pruning, which typically treats different frames with similar importance regardless of their temporal positions, VLN places greater emphasis on the current frame for immediate navigation decisions. Historical frames mainly provide supplementary navigation cues. Therefore, our method adopts different token selection strategies for current and historical frames.

  • Compared with manipulation-oriented VLA token pruning, which often involves relatively fixed viewpoints and focuses on object-state changes and robotic motion, VLN involves continuous viewpoint changes and long-horizon navigation. Thus, our method focuses more on preserving critical landmarks, capturing instruction-relevant cues that may no longer be visible, and reducing redundancy across historical observations.

  • Our VLN-specific design. We employ semantic coverage for current-frame selection and an expanded candidate pool for historical frames. Instruction-aware scoring and semi-relaxed OT further preserve task-relevant landmarks and suppress temporal redundancy.

Overall, while these approaches share broadly applicable token-pruning techniques, their technical priorities and designs differ fundamentally due to distinct task requirements.


Dear Reviewer jNBX,

Thank you again for your careful review and constructive suggestions. We sincerely appreciate the time and effort you have devoted to evaluating our work.

We have provided point-by-point responses, additional experiments, and clarifications in our rebuttal. As the discussion period draws to a close, we would be grateful if you could let us know whether these responses have addressed your concerns and which issues, if any, remain unresolved. We welcome further discussion and would be happy to provide timely clarifications before the deadline.

Thank you for your careful consideration and the effort you have invested in helping us improve our work.

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.