acceptodds
Under review as a conference paper at ICLR 2027

SGPrune: Subspace-Guided Token Pruning for Multi-View Vision-Language Models in Autonomous Driving

Abstract

Multi-view vision-language models (VLMs) support scene understanding and reasoning in autonomous driving, but processing visual tokens from multiple cameras incurs substantial computational and memory costs. Effective pruning must preserve task-relevant content and visual diversity while allocating a limited token budget across views. We propose SGPrune, a training-free framework for subspace-guided visual token pruning in multi-view VLMs. SGPrune combines three stages: Attention-Based Token Selection, PCA Subspace Construction, and Adaptive Multi-View Token Pruning. The first two stages run offline, using text-to-vision attention to select question-relevant visual tokens from a subset of the training data and summarize them into view-specific PCA subspaces. The third stage operates online during inference using the stored subspaces to estimate token relevance and guide diversity-aware selection before the multimodal projector. This selection is performed jointly across views under a shared budget, yielding input-adaptive view allocations. Experiments on autonomous driving benchmarks show that SGPrune achieves state-of-the-art performance among recent pruning methods, improving the DriveLMM-o1 Overall Reasoning score by 1.13 points and the MAPLM average score by 2.53 points over state-of-the-art baselines at about 90% token reduction. Under the same setting, SGPrune achieves a prefill speed on DriveLMM-o1 and a speed on DriveMM compared to the vanilla models, offering a superior balance between performance and efficiency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-Language Models, Visual Token Pruning, Multi-View Inference, Autonomous Driving, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_7caxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SGPrune, a new training-free method for multi-view visual language model (VLM) visual token pruning in autonomous driving. SGPrune consists of two offline processing stages and one online inference stage. The experimental results in the paper have, to some extent, verified its effectiveness and efficiency.

Critical strengths

The article is easy to understand.

This article proposes a visual token-pruning framework that requires no training.

Critical weaknesses

It seems that the innovation of this article is a simple combination of the previous pruning schemes based on attention and diversity. What is the essential difference between this article and the previous methods?

In Table 1, the method proposed in this article does not show significant performance improvement compared to the previous SOTA method. The existing ablation experiments cannot convince me of the effectiveness of the three stages of SGPrune. Suggest adding visual results to demonstrate the effectiveness of SGPrune at each stage.

Questions for authors

What is the computational complexity of this pruning method now? What about the comparison with the previous method? Do we need to consider the offline operations of the first two stages of SGPrune when comparing computational efficiency?

4 comments
AuthorsOpenReview ↗

Weakness 1: Essential difference from prior attention-based and diversity-based pruning

We thank the reviewer for this question on novelty. Your comment helped us state more precisely where SGPrune departs from prior attention-based and diversity-based methods.

SGPrune differs from prior methods in three respects.

Attention is used only offline. Attention-based pruning methods compute text-to-visual attention for every input at inference time, for example inside the decoder as in SparseVLM (Table 5, †). SGPrune uses question-conditioned attention only during offline calibration to select tokens for subspace construction. Online selection needs no decoder pass and no attention map, and it runs before the visual projector. The selection overhead is 1.2 ms on DriveLMM-o1 and 1.6 ms on DriveMM (Table 5).

Relevance is a reusable subspace, not an attention score. The attention-selected tokens are summarized by a PCA subspace for each view, and a new token is scored by its normalized projection magnitude onto that subspace. This score is not interchangeable with simpler summaries of the same tokens: replacing it with the cosine similarity or distance to their mean, the unnormalized projection magnitude, or the distance to the fitted affine subspace lowers the MAPLM score by 1.37 to 8.00 points (Table 8).

Relevance and diversity are combined in one selection over all views. The relevance score enters as a rank-based weight on the farthest-point sampling distance, and selection runs once over the pooled tokens of all views under a shared budget, so the number of tokens kept in each view is determined for each input. Prune2Drive, the closest diversity-based method, applies farthest-point sampling with per-view pruning ratios that are searched offline; its paper reports that this search converges within 3 H100 GPU hours [1]. SGPrune requires no ratio search, and its offline step takes 20 min 45 s on one H200 for DriveLMM-o1. SGPrune determines the per-view counts without any per-view ratio being searched or set in advance; under matched budgets it scores 0.01 to 0.15 points above a variant with a fixed equal allocation (Table 9).

Neither component alone reaches the combined method. At 90.1% token reduction on MAPLM, relevance-only selection scores 58.95 and diversity-only selection scores 59.38, against 67.41 for SGPrune (Table 4).

[1] Prune2Drive: A Plug-and-Play Framework for Accelerating Vision-Language Models in Autonomous Driving. CVPR 2026. arXiv:2508.13305v2, Section 3.2.

AuthorsOpenReview ↗

Weakness 2: Significance of the improvement in Table 1 and effectiveness of each stage

We thank the reviewer for this comment and for suggesting visual results for each stage. Your comment prompted us to add a significance analysis for Table 1 and a stage-wise qualitative figure, which helped us present the strengths of the paper more clearly.

Statistical significance. The revised PDF adds a paired frame-level bootstrap (10,000 resamples) of the judge scores on DriveLMM-o1 at 90.2% token reduction, for the configuration reported in Table 1 (appendix, “Statistical Significance on DriveLMM-o1”). The difference in Overall Reasoning between SGPrune and each baseline is as follows.

BaselineDifference95% CI99.5% CI
SparseVLM+5.31[4.72, 5.94][4.45, 6.18]
HAWK+2.81[2.28, 3.34][2.06, 3.58]
PruneSID+1.14[0.63, 1.65][0.41, 1.86]
ZOO-Prune+1.20[0.69, 1.71][0.47, 1.93]
Prune2Drive+1.13[0.62, 1.66][0.42, 1.87]

All intervals exclude zero, including the Bonferroni-adjusted 99.5% intervals. At this budget SGPrune scores 76.77 with 150 of 1,536 visual tokens, against 76.57 for the unpruned model.

Effectiveness of each stage. For each stage we report the end-task effect of removing or replacing it. Scores are Overall Reasoning on DriveLMM-o1 and the mean of FRM and QNS on MAPLM; percentages in the column headers are token reduction rates. Table 4 was part of the original submission; Tables 8 and 9 are new.

Stage 1, attention-based token selection. We fitted the subspaces on randomly chosen tokens instead of attention-selected tokens (Table 4).

Calibration tokensDriveLMM-o1 75.0%DriveLMM-o1 90.2%MAPLM 75.3%MAPLM 90.1%
Random76.3475.3163.8160.58
Attention-selected (SGPrune)77.1176.7769.4167.41

Attention-based selection is higher at every budget, by 0.77 and 1.46 points on DriveLMM-o1 and by 5.60 and 6.83 points on MAPLM.

Stage 2, subspace relevance scoring. Removing the relevance score (diversity-only selection) lowers the score by 0.93 and 1.22 points on DriveLMM-o1 and by 5.94 and 8.03 points on MAPLM (Table 4). Replacing it with simpler scores computed from the same tokens lowers the MAPLM score by 1.37 to 8.00 points (Table 8). On the full MAPLM test split (6,000 frame-question pairs), the 288 tokens with the highest relevance score hold 21.5% of the question's attention mass, against 9.8% for 288 random tokens, although the score does not read the question.

Stage 3, adaptive multi-view token pruning. Table 9 compares SGPrune with a variant that assigns equal token quotas to all views under the same total budget. Both use the same subspaces, relevance-weight strength and selection rule.

VariantDriveLMM-o1 75.0%DriveLMM-o1 90.2%MAPLM 75.3%MAPLM 90.1%
Fixed equal allocation76.9676.7669.2667.37
SGPrune77.1176.7769.4167.41

SGPrune is 0.01 to 0.15 points higher without any per-view quota being set in advance.

Visual results. Following your suggestion, we added a stage-wise qualitative comparison to the appendix section “Qualitative Visualization” (Figure 5, “Stage-wise qualitative comparison”). It shows (i) the tokens kept by random and by attention-based selection for the same image-question pair, (ii) the relevance scores of the same visual tokens under subspaces fitted on random and on attention-selected calibration tokens, and (iii) the tokens kept by fixed allocation with diversity-only selection and by SGPrune under the same budget of 150 tokens. The additions are highlighted in blue in the revised PDF.

AuthorsOpenReview ↗

Question 1: Computational complexity, comparison with prior methods, and offline operations

We thank the reviewer for this question. Your comment helped us report the online complexity, the measured cost and the offline cost separately.

Online complexity. The revised PDF states the online complexity in the appendix (“Efficiency: Timing Protocol, Complexity and Offline Cost”). For NN pooled visual tokens of width dd, subspace dimension cc and budget KK, SGPrune costs O(Ndc)O(Ndc) for the projections, O(Nlog⁡N)O(N\log N) for ranking, O(N2d)O(N^2 d) for the cosine-distance matrix and O(KN)O(KN) for the farthest-point steps. The vision encoder and the decoder are unchanged. Selection plus the projector on the kept tokens costs 30 GFLOPs on DriveLMM-o1 and 32 GFLOPs on DriveMM, less than the projector pass over all tokens in the unpruned model (103 and 125 GFLOPs). Counted the same way, PruneSID costs 122 and 195 GFLOPs, HAWK 195 and 262 GFLOPs, and ZOO-Prune 13.3 and 16.2 TFLOPs.

Measured comparison. The revised Table 5 reports the measured cost of every method at about 90% token reduction on one NVIDIA H200 at batch size 1, now with selection time, TTFT and estimated end-to-end latency. Below we reproduce all of its rows and columns. Latencies are in ms; parentheses give speedups over the unpruned model. Score is Overall Reasoning for DriveLMM-o1 and the mean of FRM and QNS for MAPLM.

DriveLMM-o1: 90.2% token reduction

MethodTokensSel.PrefillTTFTend-to-endFLOPs (%)KV (MB)Score
Vanilla DriveLMM-o11,536—171.5205.8 (1.00×)6694 (1.00×)100.0232.476.57
SparseVLM (ICML 2025)1503.8†55.389.5 (2.30×)6559 (1.02×)21.850.771.46
HAWK (CVPR 2026)1500.946.181.1 (2.54×)6567 (1.02×)22.050.773.96
PruneSID (ICLR 2026)15017.845.897.6 (2.11×)6678 (1.00×)21.7‾\underline{21.7}50.775.63
ZOO-Prune (CVPR 2026)15040.846.1121.1 (1.70×)6595 (1.01×)73.150.775.57
Prune2Drive (CVPR 2026)1506.0†56.791.0 (2.26×)6592 (1.02×)24.056.4‾\underline{56.4}75.64‾\underline{75.64}
SGPrune1501.2‾\underline{1.2}45.981.2‾\underline{81.2} (2.53×)6626 (1.01×)21.350.776.77

DriveMM / MAPLM: 90.1% token reduction

MethodTokensSel.PrefillTTFTend-to-endFLOPs (%)KV (MB)Score
Vanilla DriveMM2,916—301.6326.9 (1.00×)343.9 (1.00×)100.0410.472.61
SparseVLM (ICML 2025)2888.2†49.675.0 (4.36×)91.6 (3.76×)16.065.950.58
HAWK (CVPR 2026)2880.827.853.6 (6.09×)70.4 (4.88×)15.865.949.48
PruneSID (ICLR 2026)28832.027.884.9 (3.85×)101.5 (3.39×)15.7‾\underline{15.7}65.963.46
ZOO-Prune (CVPR 2026)28860.627.9113.5 (2.88×)130.2 (2.64×)50.365.963.83
Prune2Drive (CVPR 2026)2882.1†36.361.5 (5.31×)78.2 (4.40×)18.276.8‾\underline{76.8}64.88‾\underline{64.88}
SGPrune2881.6‾\underline{1.6}30.657.6‾\underline{57.6} (5.68×)74.5‾\underline{74.5} (4.62×)15.365.967.41

† Selection occurs inside the decoder and is already included in Prefill.

Time to first token (TTFT) includes the vision encoder, selection, the projector and LLM prefill; end-to-end adds decoding of the mean answer (192 and 2 tokens). SGPrune has the second lowest selection time and TTFT on both models after HAWK, and the highest score among the pruned methods.

Offline operations. The first two stages run once for a given model and camera setup and are not repeated at inference, so we report them separately from per-input latency. For DriveLMM-o1 on one NVIDIA H200, attention-based token selection takes 19 min and PCA subspace construction for all six views takes 1 min 45 s, 20 min 45 s in total. No model weights are updated. SparseVLM, HAWK, PruneSID and ZOO-Prune have no offline stage. Prune2Drive has one: its per-view ratio search is reported to converge within 3 H100 GPU hours [1]. The same appendix section also gives the timing protocol and the offline cost.

[1] Prune2Drive: A Plug-and-Play Framework for Accelerating Vision-Language Models in Autonomous Driving. CVPR 2026. arXiv:2508.13305v2, Section 3.2.

AuthorsOpenReview ↗

We Are Looking Forward to Your Reply

Thank you for your thoughtful and constructive review. In response to your comments, we clarified how SGPrune differs from prior attention-based and diversity-based methods and added the following experiments and analyses:

  • Statistical significance: A paired frame-level bootstrap analysis with 10,000 resamples to assess the improvements over the compared pruning methods on DriveLMM-o1 at 90.2% token reduction.
  • Additional ablations: Comparisons with alternative relevance scores and fixed per-view allocation under matched total token budgets (Tables 8 and 9).
  • Stage-wise visualization: Qualitative comparisons illustrating all three stages of SGPrune, following your suggestion (Figure 5).
  • Computational efficiency: Expanded latency reporting, including selection overhead, TTFT, and estimated end-to-end latency (Table 5), together with online complexity and offline preparation costs.

As the discussion period approaches its end, we would appreciate your feedback on whether these responses and additional results address your concerns. We would be happy to clarify any remaining points.

Thank you again for your time and consideration.

Best regards, The Authors

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

SGPrune builds view-specific PCA subspaces from attention-selected calibration tokens, then uses them to guide relevance-weighted diversity pruning. It keeps the model weights frozen and allocates a shared token budget across views.

Critical strengths
  • The accuracy and token-budget tradeoff looks useful on both driving models. At roughly 90% pruning, SGPrune improves over the strongest reported pruning baselines on DriveLMM-o1 and MAPLM.
  • The component ablations help explain the gains. Relevance alone and diversity alone both fall short, particularly on DriveMM. The supplement also provides calibration subsets and fitted bases.
Critical weaknesses
  • Online selection does not use the current question. The same images therefore receive the same retained tokens even for different questions. Appendix A.3 also fits a separate basis for each benchmark. It is unclear how well the learned relevance prior transfers when the question distribution changes.
  • Table 5 reports prefill speedups, while decoding speed is essentially unchanged. For the driving setting, I would also want end-to-end latency, including vision encoding, pruning, and answer generation. The prefill numbers alone do not show the overall latency benefit.

Overall: Weak acceptance. The driving results and ablations are convincing enough to make the method useful. I would keep the claims about question relevance and practical acceleration narrower until the two points above are clarified.

Questions for authors
  • How does a fixed basis perform on held-out question types or a different driving dataset without recalibration?
  • What is the end-to-end latency at the reported token budgets, with the vision encoder, selection overhead, and decoding all included?
4 comments
AuthorsOpenReview ↗

Weakness 1 & Question 1: Question-agnostic online selection and transfer of the fixed basis

We acknowledge that our online selection does not condition on the current question: given the same input images, it selects the same tokens regardless of the question. This is a deliberate design trade-off. Specifically, the subspaces are fitted using tokens selected by question-conditioned attention across calibration questions. Thus, they capture the visual information typically attended to by questions within a task, rather than adapting to any specific question at inference time.

In return, our selection operates before the visual projector and requires neither a decoder pass nor access to attention maps. Moreover, token selection can be performed once per frame, and the selected tokens can be reused across multiple questions. Accordingly, in our paper, "question-relevant" refers to relevance learned from the calibration questions, rather than relevance conditioned on the current question.

To measure how well the prior transfers, we ran three tests without recalibration on the target setting.

Held-out question types (MAPLM, DriveMM). MAPLM has four question types: scene type (SCN), data quality (QLT), lane count (LAN) and intersection (INT). We fitted subspaces on two types only and evaluated on all four. Everything else follows the paper's configuration (M=75M=75, nv=100,000n_v=100{,}000, c=64c=64, β=2\beta=2), so the calibration question types are the only difference. We report the mean of FRM and QNS; differences from the paper's configuration are paired frame-level bootstrap estimates with 95% CIs.

Calibration questions75.3%90.1%Difference from all four types (75.3% / 90.1%)
All four types (paper)69.4167.41–
LAN + INT only (SCN, QLT held out)69.4767.79+0.06 [−0.60, +0.72] / +0.38 [−0.47, +1.23]
SCN + QLT only (LAN, INT held out)68.1965.19−1.22 [−1.91, −0.53] / −2.22 [−3.17, −1.26]

Transfer is asymmetric. Subspaces fitted only on scene and quality questions reduce the overall mean score by 1.22 and 2.22 points at 75.3% and 90.1% token reduction, respectively. At 90.1% reduction, held-out lane-count accuracy falls from 73.80% to 70.53%.

Different driving dataset (nuScenes to MAPLM). We evaluated MAPLM with subspaces fitted on nuScenes, which differs in location, camera rig and question set; MAPLM's BEV point-cloud view has no nuScenes counterpart and receives a subspace pooled over all cameras. The scores are 66.58 and 64.12 at 75.3% and 90.1% reduction, 2.83 and 3.29 points below the in-domain subspaces (69.41 and 67.41). At 90.1% reduction this is on par with ZOO-Prune (63.83) and PruneSID (63.46); at 75.3% it is below them (67.84 and 67.95).

Unseen conditions (DriveLMM-o1). The calibration subset contains no night scenes. On the 409 night-time test items at 90.2% reduction, SGPrune scores 79.45 against 77.00 (Prune2Drive), 77.00 (PruneSID) and 77.37 (ZOO-Prune), with gains of 2.1 to 2.4 points whose 95% CIs lie above zero. At 75.0% reduction the differences are not significant.

These results suggest that transfer depends on which question types are represented in calibration and on the target setting. SGPrune is intended for settings in which the camera setup and the question distribution are known in advance, and the subspaces are fitted for that setting. Recalibration is a one-time offline step (about 20 minutes on one GPU for DriveLMM-o1), so we recommend it when the question distribution or the camera setup changes.

AuthorsOpenReview ↗

Weakness 2 & Question 2: End-to-end latency

We sincerely thank the reviewer for pointing out the need for greater clarity regarding our latency measurements.

In the original submission, visual encoding and selection overhead were already included in the reported prefill timing, along with visual projection and LLM prefill. However, this measurement scope was not explicitly clarified. To address this ambiguity, we have revised the PDF to explicitly label this quantity as Time to First Token (TTFT) and expanded Table 5 to report absolute latency, selection overhead, and estimated end-to-end speedups.

Specifically, TTFT is computed as the sum of separately measured stage medians, including visual encoding, selection, projection, and LLM prefill, measured on a single H200 GPU with batch size 1.

Below, we provide the complete updated Table 5, including all methods and metrics, for both DriveLMM-o1 and MAPLM at approximately 90% token reduction. The same updated table is also included in the revised PDF (Table 5) for reference. Latencies are reported in milliseconds, with speedups relative to the corresponding vanilla model in parentheses. The score represents Overall Reasoning for DriveLMM-o1 and (FRM + QNS)/2 for MAPLM.

DriveLMM-o1: 90.2% token reduction

MethodTokensSel.PrefillTTFTend-to-endFLOPs (%)KV (MB)Score
Vanilla DriveLMM-o11,536—171.5205.8 (1.00×)6694 (1.00×)100.0232.476.57
SparseVLM (ICML 2025)1503.8†55.389.5 (2.30×)6559 (1.02×)21.850.771.46
HAWK (CVPR 2026)1500.946.181.1 (2.54×)6567 (1.02×)22.050.773.96
PruneSID (ICLR 2026)15017.845.897.6 (2.11×)6678 (1.00×)21.7‾\underline{21.7}50.775.63
ZOO-Prune (CVPR 2026)15040.846.1121.1 (1.70×)6595 (1.01×)73.150.775.57
Prune2Drive (CVPR 2026)1506.0†56.791.0 (2.26×)6592 (1.02×)24.056.4‾\underline{56.4}75.64‾\underline{75.64}
SGPrune1501.2‾\underline{1.2}45.981.2‾\underline{81.2} (2.53×)6626 (1.01×)21.350.776.77

DriveMM / MAPLM: 90.1% token reduction

MethodTokensSel.PrefillTTFTend-to-endFLOPs (%)KV (MB)Score
Vanilla DriveMM2,916—301.6326.9 (1.00×)343.9 (1.00×)100.0410.472.61
SparseVLM (ICML 2025)2888.2†49.675.0 (4.36×)91.6 (3.76×)16.065.950.58
HAWK (CVPR 2026)2880.827.853.6 (6.09×)70.4 (4.88×)15.865.949.48
PruneSID (ICLR 2026)28832.027.884.9 (3.85×)101.5 (3.39×)15.7‾\underline{15.7}65.963.46
ZOO-Prune (CVPR 2026)28860.627.9113.5 (2.88×)130.2 (2.64×)50.365.963.83
Prune2Drive (CVPR 2026)2882.1†36.361.5 (5.31×)78.2 (4.40×)18.276.8‾\underline{76.8}64.88‾\underline{64.88}
SGPrune2881.6‾\underline{1.6}30.657.6‾\underline{57.6} (5.68×)74.5‾\underline{74.5} (4.62×)15.365.967.41

† Selection occurs inside the decoder and is already included in Prefill.

SGPrune incurs only 1.2 ms of selection overhead on DriveLMM-o1 and 1.6 ms on DriveMM, both included in the reported TTFT. We further estimate end-to-end latency as TTTFT+(L−1)tdecT_{\mathrm{TTFT}}+(L-1)t_{\mathrm{dec}}, using the measured decoding cost with L=192L=192 and L=2L=2 for DriveLMM-o1 and DriveMM, respectively. The relatively long answer generation for DriveLMM-o1 limits the overall end-to-end speedup despite the substantial reduction in TTFT.

Total decoding time also depends strongly on the number of generated output tokens. Our primary goal is to reduce LLM prefill latency by pruning visual input tokens. Further reducing end-to-end latency in driving applications, including the cost of decoding long answers, remains an important direction for future work.

Image loading, preprocessing, and tokenization are excluded from these measurements.

Further details on the timing protocol, measurement configurations, online computational complexity, and offline preparation cost are provided in the appendix, "Efficiency: Timing Protocol, Complexity and Offline Cost."

Thanks. The asymmetric results help explain where the calibration prior works and where recalibration is needed. The narrower interpretation of “question-relevant” also addresses my concern. good!

Thank you for your positive feedback and for acknowledging our clarifications.

We are glad that the additional transfer experiments and the more precise interpretation of “question-relevant” have addressed your concerns about the scope and transferability of the calibration prior. We also hope that our expanded latency analysis, including selection overhead and estimated end-to-end latency, has clarified both the practical efficiency benefits of SGPrune and their dependence on answer length.

Given these additional experiments and clarifications, we would appreciate it if you could reconsider your overall assessment of the paper.

Thank you again for your thoughtful review and constructive feedback.

Reviewer_TB6yOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SGPrune, a training-free token pruning method for multi-view vision-language models. It constructs view-specific PCA subspaces from attention-selected tokens offline, then combines subspace relevance with diversity to select tokens across views under a shared budget. Experiments cover two driving models and seven general vision-language benchmarks, reporting improved performance over pruning baselines and prefill acceleration at approximately 90% token reduction.

Critical strengths
  1. Practical design. Offline relevance priors enable pruning before the multimodal projector without extracting decoder attention online. The supplementary material includes core code and precomputed subspaces.
  2. Meaningful component ablations. Table 4 supports the complementary roles of relevance and diversity: on DriveMM at approximately 90% pruning, the full method achieves 67.41, compared with 58.95 for relevance-only and 59.38 for diversity-only selection.
Critical weaknesses
  1. Inconsistent MAPLM baseline metrics. Table 2 reports FRM=59.07% and QNS=96.15% for the unpruned model. With four questions per frame and FRM requiring all four answers to be correct, these metrics must satisfy QNS≤(3+FRM)/4QNS\leq(3+FRM)/4, expressed as fractions. The reported FRM therefore permits a QNS of at most 89.77%. This discrepancy requires clarification, including its implications for Table 5.
  2. The independent benefit of adaptive view allocation is unverified. The appendix provides only qualitative comparisons. Existing ablations assess relevance, diversity, and view-specific subspaces, but do not quantify fixed versus adaptive allocation under otherwise matched settings.
Questions for authors
  1. Does the FRM/QNS inconsistency in Table 2 arise from a reporting error or different evaluation protocols? Does consistent recomputation affect the conclusions?
  2. How much does adaptive allocation improve over fixed quotas when the subspaces, selection criteria, and total token budget are held constant?
  3. Were hyperparameters selected using an independent validation set? Is the 1.13-point improvement on DriveLMM-o1 statistically reliable?
  4. Does prefill timing include visual encoding and pruning overhead? Please provide absolute latency and end-to-end speedups.
8 comments
AuthorsOpenReview ↗

Weakness 1 & Question 1: MAPLM metrics

We thank the reviewer for the constructive feedback. Your comments helped us identify a reporting error and re-examine our results and procedures. We appreciate this opportunity to improve the accuracy, transparency, and quality of our work.

We also provide a revised PDF with changes highlighted in blue.

We apologize for this reporting error. When manually entering the results into the table, we mistakenly recorded QNS as 96.15% instead of 86.15%, which also led to the incorrect average of 77.61. The corrected values are:

FRM (%)QNS (%)Average
59.0786.1572.61

These satisfy 0.8615≤(3+0.5907)/4=0.8976750.8615 \leq (3+0.5907)/4=0.897675. The evaluation protocol and metric definitions are unchanged. We have corrected the unpruned DriveMM entries in Tables 2 and 5. This transcription correction does not change the scores or relative rankings of SGPrune and the other pruning methods.

We thank the reviewer for identifying this discrepancy. Your comment also prompted us to recheck the other reported values and calculations, providing an opportunity to improve the accuracy of the manuscript.

AuthorsOpenReview ↗

Question 3(a): Hyperparameter selection and sensitivity

We sincerely thank the reviewer for this important question and the opportunity to clarify our hyperparameter selection procedure.

Yes, all hyperparameters were selected solely based on mean performance on an independent validation set across the two token budgets, without using or referencing test results during selection.

For completeness and transparency, we provide the full validation results alongside the corresponding test scores below. Notably, SGPrune consistently outperforms the compared pruning methods across both token budgets for DriveLMM-o1 and MAPLM, even when hyperparameter settings are varied, indicating that its performance advantage is not limited to a particular configuration.

Further details on the validation set are provided in Appendix A.3 (Additional Ablation Studies).

  1. DriveLMM-o1 — Subspace Construction | M | nvn_v | c | Val. 75.0% | Val. 90.2% | Val. Mean | Test 75.0% | Test 90.2% | Test Mean | |---:|---:|---:|---:|---:|---:|---:|---:|---:| | Varying M ||||||||| | 25 | 100,000 | 64 | 79.00 | 79.37 | 79.19 | 76.91 | 76.80 | 76.86 | | 50 | 100,000 | 64 | 79.66 | 78.76 | 79.21 | 77.11 | 76.77 | 76.94 | | 75 | 100,000 | 64 | 79.81 | 78.56 | 79.19 | 76.94 | 76.49 | 76.72 | | Varying nvn_v ||||||||| | 50 | 50,000 | 64 | 78.28 | 78.90 | 78.59 | 76.96 | 76.73 | 76.85 | | 50 | 100,000 | 64 | 79.66 | 78.76 | 79.21 | 77.11 | 76.77 | 76.94 | | 50 | 150,000 | 64 | 79.27 | 78.27 | 78.77 | 76.99 | 76.86 | 76.93 | | Varying c ||||||||| | 50 | 100,000 | 32 | 78.74 | 78.55 | 78.65 | 77.09 | 76.74 | 76.92 | | 50 | 100,000 | 64 | 79.66 | 78.76 | 79.21 | 77.11 | 76.77 | 76.94 | | 50 | 100,000 | 128 | 78.78 | 78.36 | 78.57 | 76.58 | 76.30 | 76.44 |

  2. MAPLM using DriveMM — Subspace Construction | M | nvn_v | c | Val. 75.3% | Val. 90.1% | Val. Mean | Test 75.3% | Test 90.1% | Test Mean | |---:|---:|---:|---:|---:|---:|---:|---:|---:| | Varying M ||||||||| | 50 | 100,000 | 64 | 68.23 | 65.85 | 67.04 | 69.32 | 67.27 | 68.30 | | 75 | 100,000 | 64 | 68.45 | 66.04 | 67.25 | 69.41 | 67.41 | 68.41 | | 100 | 100,000 | 64 | 68.60 | 65.25 | 66.93 | 68.83 | 66.12 | 67.48 | | Varying nvn_v ||||||||| | 75 | 50,000 | 64 | 67.95 | 65.28 | 66.62 | 68.87 | 67.04 | 67.96 | | 75 | 100,000 | 64 | 68.45 | 66.04 | 67.25 | 69.41 | 67.41 | 68.41 | | 75 | 150,000 | 64 | 68.95 | 65.42 | 67.19 | 68.94 | 66.73 | 67.84 | | Varying c ||||||||| | 75 | 100,000 | 32 | 67.21 | 65.28 | 66.25 | 68.19 | 66.13 | 67.16 | | 75 | 100,000 | 64 | 68.45 | 66.04 | 67.25 | 69.41 | 67.41 | 68.41 | | 75 | 100,000 | 128 | 68.63 | 65.84 | 67.24 | 68.86 | 66.83 | 67.85 |

  3. InternVL3-8B — Subspace Construction | M | nvn_v | c | Val. 75.0% | Val. 90.2% | Val. Mean | Test 75.0% | Test 90.2% | Test Mean | |---:|---:|---:|---:|---:|---:|---:|---:|---:| | Varying M ||||||||| | 10 | 100,000 | 128 | 79.91 | 63.12 | 71.52 | 82.70 | 70.82 | 76.76 | | 25 | 100,000 | 128 | 79.94 | 63.66 | 71.80 | 83.01 | 71.17 | 77.09 | | 50 | 100,000 | 128 | 79.43 | 63.08 | 71.26 | 82.72 | 70.46 | 76.59 | | Varying nvn_v ||||||||| | 25 | 50,000 | 128 | 79.55 | 63.90 | 71.73 | 82.60 | 71.15 | 76.88 | | 25 | 100,000 | 128 | 79.94 | 63.66 | 71.80 | 83.01 | 71.17 | 77.09 | | 25 | 150,000 | 128 | 79.59 | 63.79 | 71.69 | 82.80 | 70.96 | 76.88 | | Varying c ||||||||| | 25 | 100,000 | 64 | 78.81 | 62.56 | 70.69 | 82.26 | 70.51 | 76.39 | | 25 | 100,000 | 128 | 79.94 | 63.66 | 71.80 | 83.01 | 71.17 | 77.09 | | 25 | 100,000 | 256 | 80.13 | 63.45 | 71.79 | 83.52 | 70.66 | 77.09 |

  4. DriveLMM-o1 — Relevance-weight Strength | β\beta | Val. 75.0% | Val. 90.2% | Val. Mean | Test 75.0% | Test 90.2% | Test Mean | |---:|---:|---:|---:|---:|---:|---:| | 1 | 79.16 | 78.50 | 78.83 | 76.87 | 76.29 | 76.58 | | 2 | 79.66 | 78.76 | 79.21 | 77.11 | 76.77 | 76.94 | | 3 | 79.29 | 78.84 | 79.07 | 77.06 | 76.55 | 76.81 |

  5. MAPLM using DriveMM — Relevance-weight Strength | β\beta | Val. 75.3% | Val. 90.1% | Val. Mean | Test 75.3% | Test 90.1% | Test Mean | |---:|---:|---:|---:|---:|---:|---:| | 1 | 68.14 | 64.95 | 66.55 | 68.13 | 64.71 | 66.42 | | 2 | 68.45 | 66.04 | 67.25 | 69.41 | 67.41 | 68.41 | | 3 | 67.44 | 65.17 | 66.31 | 68.57 | 66.25 | 67.41 |

  6. InternVL3-8B — Relevance-weight Strength | β\beta | Val. 75.0% | Val. 90.2% | Val. Mean | Test 75.0% | Test 90.2% | Test Mean | |---:|---:|---:|---:|---:|---:|---:| | 0.25 | 78.43 | 62.00 | 70.22 | 81.94 | 69.90 | 75.92 | | 0.5 | 79.94 | 63.66 | 71.80 | 83.01 | 71.17 | 77.09 | | 1 | 79.44 | 63.33 | 71.39 | 82.91 | 70.64 | 76.78 |

We also provide a revised PDF with changes highlighted in blue.

AuthorsOpenReview ↗

Question 3(b): Statistical reliability

We sincerely thank the reviewer for raising this important question regarding the statistical reliability of our reported improvement. To address this concern, we have added a paired frame-level bootstrap analysis for DriveLMM-o1 at 90.2% token reduction, using the configuration reported in Table 1 (M=50M=50, nv=100,000n_v=100{,}000, c=64c=64, β=2\beta=2).

Specifically, we draw 10,000 bootstrap replicates, each resampling the 539 test frames with replacement while keeping all associated questions (4,634 QA pairs in total) and the matched scores of all methods together. For each replicate, we compute the mean Overall Reasoning difference between SGPrune and each baseline. We report both 95% percentile confidence intervals (CIs) and Bonferroni-adjusted 99.5% CIs, accounting for the ten evaluated comparisons (five baselines across two token budgets, m=10m=10).

BaselineΔ\Delta95% CI99.5% CI
SparseVLM+5.31[4.72, 5.94][4.45, 6.18]
HAWK+2.81[2.28, 3.34][2.06, 3.58]
PruneSID+1.14[0.63, 1.65][0.41, 1.86]
ZOO-Prune+1.20[0.69, 1.71][0.47, 1.93]
Prune2Drive+1.13[0.62, 1.66][0.42, 1.87]

Importantly, all five Bonferroni-adjusted confidence intervals lie entirely above zero, indicating that the improvements over all compared pruning methods at 90.2% token reduction remain statistically significant after multiple-comparison correction. In particular, the 1.13-point improvement over Prune2Drive has an adjusted 99.5% CI of [0.42,1.87][0.42, 1.87], supporting the statistical reliability of this improvement.

These additional results are included in the revised PDF in the new appendix, "Statistical Significance on DriveLMM-o1," along with a corresponding statement added to Section 4.1.

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

This paper proposes SGPrune, a training-free visual token pruning framework for multi-view VLMs in autonomous driving. Offline, text-to-vision attention selects relevant visual tokens from calibration data, and view-specific PCA subspaces summarize their feature variation. Online, normalized subspace projection scores provide relevance weights for global diversity-aware token selection under a shared multi-view budget. Experiments demonstrate improvements over the compared pruning methods on two driving benchmarks, with additional evaluations on general vision-language benchmarks.

Critical strengths
  1. Using an offline-constructed subspace as a computationally inexpensive proxy for token relevance is an interesting design choice. It allows task-informed online pruning through simple projection operations, without extracting decoder attention during inference.

  2. The relevance-weighted diversity selection provides a balanced way to preserve important content while maintaining broad feature coverage. Performing selection jointly across views addresses redundancy both within and across cameras and naturally produces input-dependent view allocations.

Critical weaknesses
  1. PCA captures high-variance directions among attention-selected features, but these directions do not necessarily correspond to the most task-critical information, particularly for rare objects or events. The random-selection ablation supports attention-based calibration, but does not isolate the value of PCA relative to simpler alternatives. The connection between PCA projection and task relevance needs stronger justification.

  2. It is unclear how well the offline subspaces work when the input distribution changes. Each benchmark uses its own calibration data, so the current experiments do not show whether the same subspaces remain useful across different datasets, weather conditions, or camera setups.

  3. Table 5 reports promising prefill speedups, but it is unclear whether the measurements include visual encoding and the full pruning process. Reporting absolute latency, pruning overhead, and the timing setup would make the practical gains easier to assess.

Questions for authors
  1. Could you report time to first token (TTFT), including visual encoding, token pruning, and LLM prefill, to show the overall latency improvement?

  2. How does SGPrune handle distribution differences between the offline calibration data and online inputs? Would the existing subspaces remain effective, or would they need to be recalibrated?

6 comments
AuthorsOpenReview ↗

Weakness 2 & Question 2: Distribution shift between calibration data and online inputs

We thank the reviewer for raising the question of distribution shift. Your comment prompted us to evaluate the subspaces without recalibration under a different dataset and camera setup, an unseen imaging condition, and a different question distribution. These experiments allowed us to explicitly assess how performance changes across these settings and better understand the sensitivity of the fixed subspaces to distribution shifts.

SGPrune does not adapt the subspaces online; they are fixed after calibration. This is a deliberate design trade-off. Specifically, the subspaces are fitted using tokens selected by question-conditioned attention across calibration questions. Thus, they capture the visual information typically attended to by questions within a task, rather than adapting to any specific question at inference time.

This allows online selection before the visual projector without an additional decoder pass or attention-map computation. However, the fixed prior may be less effective when new inputs or questions require information that is poorly represented in calibration. To measure how well the subspaces transfer under such shifts, we ran three tests without recalibration on the target setting.

Different dataset and camera setup (nuScenes to MAPLM). We evaluated MAPLM with subspaces fitted on nuScenes, which differs in location, camera rig and question set. MAPLM's three cameras use the subspace of the matching nuScenes camera; its BEV point-cloud view has no matching camera, so it uses one subspace fitted on tokens from all six nuScenes cameras. The scores are 66.58 and 64.12 at 75.3% and 90.1% reduction, 2.83 and 3.29 points below the in-domain subspaces (69.41 and 67.41). At 90.1% reduction this is on par with ZOO-Prune (63.83) and PruneSID (63.46); at 75.3% it is below them (67.84 and 67.95).

Unseen imaging condition (DriveLMM-o1). The calibration subset contains no night scenes. On the 409 night-time test items at 90.2% reduction, SGPrune scores 79.45 against 77.00 (Prune2Drive), 77.00 (PruneSID) and 77.37 (ZOO-Prune), with gains of 2.1 to 2.4 points whose 95% CIs lie above zero. At 75.0% reduction, SGPrune is on par with these methods (77.74 against 76.88 to 78.20; differences not significant). We did not isolate transfer to an unseen weather condition, since rainy scenes already account for 11% of the calibration subset.

Different question distribution (MAPLM). MAPLM has four question types: scene type (SCN), data quality (QLT), lane count (LAN), and intersection (INT). We fitted subspaces on two types only and evaluated on all four, keeping M=75M=75, nv=100,000n_v=100{,}000, c=64c=64, and β=2\beta=2 fixed. We report the mean of FRM and QNS across all four types. Differences from the all-type configuration are accompanied by paired frame-level bootstrap 95% confidence intervals.

Calibration questions75.3%90.1%Difference from all four types (75.3% / 90.1%)
All four types (paper)69.4167.41–
LAN + INT only (SCN, QLT held out)69.4767.79+0.06 [−0.60, +0.72] / +0.38 [−0.47, +1.23]
SCN + QLT only (LAN, INT held out)68.1965.19−1.22 [−1.91, −0.53] / −2.22 [−3.17, −1.26]

The columns denote token reduction rates. Calibrating only on SCN and QLT reduces the overall mean score by 1.22 and 2.22 points, respectively, whereas the differences for LAN and INT calibration are not statistically significant.

In summary, the existing subspaces remain effective under a change of imaging condition, lose 1.2 to 2.2 points when the calibration questions do not cover the content that new questions need, and lose about 3 points when the dataset and camera setup change. For a new dataset or camera setup we therefore recommend recalibration, which is a one-time offline step (about 20 minutes on one GPU for DriveLMM-o1).

AuthorsOpenReview ↗

Weakness 1: Why PCA, and its connection to task relevance

We thank the reviewer for this question on the role of PCA. Your comment prompted us to compare PCA directly against simpler alternatives and to measure how the resulting relevance score relates to decoder attention.

Role of PCA. In SGPrune, the relevance signal comes from the calibration tokens rather than from PCA: they are selected by question-conditioned text-to-visual attention over the calibration questions. PCA summarizes these tokens as a compact subspace per view, so that a new token can be scored before the visual projector without a decoder pass or an attention map. Whether this summary retains the relevance signal is an empirical question, which we address below.

PCA against simpler alternatives. To isolate the value of PCA, we replaced the relevance score with simpler summaries of the same attention-selected tokens. Everything else follows the paper's configuration on MAPLM (same calibration tokens, M=75M=75, nv=100,000n_v=100{,}000, c=64c=64, β=2\beta=2, same diversity term and allocation). We report the mean of FRM and QNS. Point differences are calculated from the displayed scores, while the 95% CIs are obtained by paired frame-level bootstrap using unrounded scores.

Relevance score75.3%90.1%Difference from SGPrune (75.3% / 90.1%)
Cosine similarity to the mean μv\boldsymbol\mu_v of the selected tokens65.6162.27−3.80 [−4.97, −2.62] / −5.14 [−6.49, −3.79]
Distance to the mean μv\boldsymbol\mu_v64.1059.41−5.31 [−6.60, −4.02] / −8.00 [−9.47, −6.56]
Unnormalized projection magnitude ∥Uvf∥2\lVert\mathbf U_v\mathbf f\rVert_265.1661.21−4.25 [−5.50, −3.00] / −6.20 [−7.61, −4.81]
Distance to the fitted affine subspace μv+span(Uv)\boldsymbol\mu_v+\mathrm{span}(\mathbf U_v)68.0465.82−1.37 [−2.34, −0.39] / −1.59 [−2.74, −0.45]
Normalized projection magnitude (SGPrune)69.4167.41–

The columns denote token reduction rates. Every alternative is significantly below SGPrune at both reduction rates. The two scores that use only the mean of the selected tokens, without PCA, lose 3.8 to 8.0 points. The two scores that use the same PCA subspace in a different way are also lower: omitting the normalization by the token norm loses 4.3 to 6.2 points, and replacing the projection magnitude with the distance to the subspace loses 1.4 to 1.6 points. We have added this comparison to the appendix (“Effect of the relevance score”, Table 8); the added paragraph and table are highlighted in blue in the revised PDF.

Connection to task relevance. We measured directly whether the question-free score ranks tokens as the decoder's question-conditioned attention does. On the full MAPLM test split (6,000 frame-question pairs), the 288 tokens with the highest SGPrune relevance hold 21.5% of the question's attention mass, against 9.8% for a random choice of 288 tokens and 15.4% when the subspaces are fitted on random patches. The subspace therefore carries the relevance signal of the attention-selected tokens to inputs it has not seen.

Rare objects and events. SGPrune uses relevance as a soft preference during diversity-based selection. For β=2\beta=2, the rank-based relevance weight ranges from e−1e^{-1} to e1e^{1}, so a low relevance score alone does not exclude a token from the candidate pool. Farthest-point sampling also rewards feature diversity, allowing tokens with distinct features to remain competitive despite low relevance weights. Removing the diversity term lowers overall MAPLM performance by 3.7 and 8.5 points at 75.3% and 90.1% token reduction, respectively (Table 4). Together with the relevance-score ablations, these results support combining a calibration-informed relevance prior with diversity-based selection in the evaluated settings.

AuthorsOpenReview ↗

Weakness 3 & Question 1: Time to first token

We thank the reviewer for requesting clearer latency reporting. Your comment helped us make explicit what the reported times include and report absolute latency, pruning overhead, and the timing setup.

Yes. SGPrune's previously reported prefill speedups include visual encoding, token selection, projection, and LLM prefill. In the revised PDF, Table 5 labels this quantity time to first token (TTFT), separates it from LLM prefill, and reports absolute latency and selection overhead for every method. TTFT is the sum of separately measured stage medians.

Vision encoding (about 34 ms on DriveLMM-o1 and 25 ms on DriveMM) and projection are included in TTFT, although they are not shown as separate columns. Below we reproduce all rows and columns of the revised Table 5. Latencies are in ms; parentheses give speedups over the unpruned model. Score is Overall Reasoning for DriveLMM-o1 and the mean of FRM and QNS for MAPLM.

DriveLMM-o1: 90.2% token reduction

MethodTokensSel.PrefillTTFTend-to-endFLOPs (%)KV (MB)Score
Vanilla DriveLMM-o11,536—171.5205.8 (1.00×)6694 (1.00×)100.0232.476.57
SparseVLM (ICML 2025)1503.8†55.389.5 (2.30×)6559 (1.02×)21.850.771.46
HAWK (CVPR 2026)1500.946.181.1 (2.54×)6567 (1.02×)22.050.773.96
PruneSID (ICLR 2026)15017.845.897.6 (2.11×)6678 (1.00×)21.7‾\underline{21.7}50.775.63
ZOO-Prune (CVPR 2026)15040.846.1121.1 (1.70×)6595 (1.01×)73.150.775.57
Prune2Drive (CVPR 2026)1506.0†56.791.0 (2.26×)6592 (1.02×)24.056.4‾\underline{56.4}75.64‾\underline{75.64}
SGPrune1501.2‾\underline{1.2}45.981.2‾\underline{81.2} (2.53×)6626 (1.01×)21.350.776.77

DriveMM / MAPLM: 90.1% token reduction

MethodTokensSel.PrefillTTFTend-to-endFLOPs (%)KV (MB)Score
Vanilla DriveMM2,916—301.6326.9 (1.00×)343.9 (1.00×)100.0410.472.61
SparseVLM (ICML 2025)2888.2†49.675.0 (4.36×)91.6 (3.76×)16.065.950.58
HAWK (CVPR 2026)2880.827.853.6 (6.09×)70.4 (4.88×)15.865.949.48
PruneSID (ICLR 2026)28832.027.884.9 (3.85×)101.5 (3.39×)15.7‾\underline{15.7}65.963.46
ZOO-Prune (CVPR 2026)28860.627.9113.5 (2.88×)130.2 (2.64×)50.365.963.83
Prune2Drive (CVPR 2026)2882.1†36.361.5 (5.31×)78.2 (4.40×)18.276.8‾\underline{76.8}64.88‾\underline{64.88}
SGPrune2881.6‾\underline{1.6}30.657.6‾\underline{57.6} (5.68×)74.5‾\underline{74.5} (4.62×)15.365.967.41

† Selection occurs inside the decoder and is already included in Prefill.

End-to-end latency is estimated as TTTFT+(L−1)tdecT_{\mathrm{TTFT}}+(L-1)t_{\mathrm{dec}}, where tdect_{\mathrm{dec}} is the measured per-token decoding latency and L=192L=192 for DriveLMM-o1 and L=2L=2 for DriveMM. The first output token is included in TTFT.

Overall latency improvement. Including visual encoding and token selection, SGPrune reduces TTFT from 205.8 to 81.2 ms on DriveLMM-o1 (2.53×) and from 326.9 to 57.6 ms on DriveMM (5.68×). The gain comes from LLM prefill, which falls from 171.5 to 45.9 ms and from 301.6 to 30.6 ms; vision encoding is unchanged, and selection adds 1.2 and 1.6 ms. The end-to-end gain depends on answer length. With decoding included, the estimated latency falls from 343.9 to 74.5 ms on DriveMM (4.62×), where answers are short, and from 6,694 to 6,626 ms on DriveLMM-o1 (1.01×), where decoding the 192-token answers dominates.

SGPrune's selection overhead is 1.2 ms on DriveLMM-o1 and 1.6 ms on DriveMM, the second lowest after HAWK (0.9 and 0.8 ms).

Timing setup. We use one NVIDIA H200, bfloat16, batch size 1, and eager attention, with 20 inputs and five timed repeats after warm-up. We take the median over repeats for each input, then over inputs, with method order shuffled. Image loading, preprocessing, transfers, tokenization, and untimed sequence assembly are excluded. The appendix, “Efficiency: Timing Protocol, Complexity and Offline Cost,” gives further details.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank You for Your Time and Consideration

We would like to sincerely thank all the reviewers and the Area Chair for taking the time and effort to carefully evaluate our work and provide valuable feedback.

In particular, we made considerable efforts to address Reviewer 7cax's concerns through detailed responses and additional experiments, including statistical significance analysis, further ablation studies, stage-wise visualizations, and expanded computational efficiency evaluations.

Although we have not received further feedback from Reviewer 7cax, we sincerely hope that our responses and additional results have adequately addressed their concerns, should they have had the opportunity to review them.

We would also like to express our sincere appreciation to the Area Chair for overseeing the discussion and evaluation process.

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.