acceptodds
Withdrawn from ICLR 2027

WaLKPruner: Cognitively Inspired Visual Token Pruning for Spatial Understanding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visual token pruning reduces the computational cost of spatial understanding in multimodal large language models. Multi-view inputs contain repeated scene content, but viewpoint changes reveal complementary observations needed to distinguish spatial relationships. Objects overlapping in one view may be separated in another. Feature-guided pruning can preserve the main scene content while discarding spatial details and contextual cues. Geometry-guided pruning maintains diversity across voxels, yet may concentrate tokens in some viewpoints while overcompressing observations from others. Effective pruning must retain representative scene content together with complementary cues along the observation path. Inspired by non-selective global perception and selective processing in human spatial exploration, we introduce WaLKPruner, a training-free framework following the principle of Walk and Look for Key Anchors. Scene abstraction and viewpoint tracking jointly guide token selection, maintaining a compact scene overview while incorporating complementary observations along the ordered observation path. Constrained selective updates limit changes to the initial scene representation while maintaining a fixed token budget. Experiments on three benchmarks show strong performance across token budgets. Efficiency analyses indicate a favorable balance between computational cost and spatial understanding. Quantitative and qualitative analyses reveal balanced token allocation across viewpoints, while ablations support combining scene and path information and retaining observation order. Code and scripts will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
cognitively inspired pruning, scene abstraction, path-guided refinement, spatial understanding

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_ryQtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes WaLKPruner, a training-free and question-independent visual token pruning method for spatial understanding. It first selects representative tokens using a feature-based coverage objective, then incorporates ordered camera motion estimated by the backbone’s VGGT geometry encoder. A constrained replacement procedure introduces trajectory-guided candidates while maintaining a fixed token budget and limiting changes to the initial scene representation. Experiments on VSI-Bench, ScanQA, and SQA3D use a Spatial-MLLM 3B backbone.

Critical strengths
  1. A practical and reasonably motivated design. Preserving complementary observations during spatial token pruning is a relevant problem. The method requires no additional training and reuses camera poses already available in the backbone. Its conservative replacement constraints provide a reasonable mechanism for incorporating trajectory information without substantially disrupting the initial selection.
  2. Promising results on the evaluated backbone. The VSI-Bench results show effective compression: retaining 25% of visual tokens achieves 50.25 compared with the reported unpruned score of 50.39. The improvements over Geo3DPruner increase under stronger compression. The ablations also provide initial evidence that scene and viewpoint information are complementary.
Critical weaknesses
  1. The incremental contribution over closely related work is insufficiently established. Representative coverage is already central to MMTok, while CoverPrune explicitly combines feature, spatial, and temporal information for 3D token compression. Motion-aware pruning has also been explored in EgoPrune. These methods do not make the proposed ordered-trajectory formulation redundant, but they substantially narrow its novelty. The paper needs clearer positioning and stronger evidence that its specific trajectory mechanism provides meaningful benefits beyond existing coverage strategies and simple view-balanced alternatives.
  2. The generality and reliability of the trajectory-guided gains remain unclear. All experiments use a single Spatial-MLLM 3B backbone and concentrate on indoor scanning benchmarks. This does not establish whether the benefits persist with stronger models or different geometric representations. Moreover, estimated camera motion directly influences token replacement, yet the paper provides no targeted evaluation of sensitivity to pose errors or failure cases. The scene-coverage constraint does not directly guarantee preservation of task-critical evidence. These omissions leave uncertainty about the method’s applicability beyond the evaluated setting.

[1] MMTok: Multimodal Coverage Maximization for Efficient Inference of VLMs

[2] CoverPrune: Coverage-Driven Token Pruning for 3D VLMs via Optimal Transport

[3] EgoPrune: Efficient Token Pruning for Egomotion Video Reasoning in Embodied Agent

Questions for authors

Can the authors clarify the distinction from CoverPrune, MMTok, and motion-aware pruning methods? In particular, how much does ordered camera motion improve over scene-only selection with equal per-view budgets or a simple temporal-coverage constraint?

Do the gains persist on a second, preferably larger, spatial-understanding backbone? Such evidence would help distinguish a generally useful selection principle from sensitivity specific to the current 3B model.

How sensitive is the selector to inaccurate camera poses? An informative test would keep the fused visual tokens fixed while perturbing only the poses supplied to the selector, comparing the resulting degradation with scene-only and view-balanced selection.

How were the baselines adapted to Spatial-MLLM?

Reviewer_u5XJOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduce WaLKPruner, a training-free framework following the principle of Walk and Look for Key Anchors. Based on Spatial-MLLM, this paper selects representative fused visual–geometric tokens using a facility-location objective. Different with other method, it constructs a motion-aware coordinate along the ordered camera trajectory and uses scene–path coverage to select tokens. It proves its efficiency on three benchmarks.

Critical strengths
  1. First to use the camera trajectory to help prune tokens in spatial VLM.
  2. Requires no additional training and the selected tokens are not related to the given questions. It means this method only remove redundant and similar tokens.
  3. The experiments demonstrate improvement over other methods.
Critical weaknesses
  1. lack of novety, and the contribution beyond existing methods are not enough.
Questions for authors
  1. it's a little weird to use a accumulated changes of the defined camera changes. Since the author use the weighted sum of translation and ratation, it does not sound reasonable.
  2. The margin from 25% to 50% is very small, compared to other method. Could you explain this phenomenon in detail? For example, prove the extra tokens from 25% to 50% are redundant?
  3. Preservation of positional information is underspecified, e.g. position IDs. I wonder how the author process the position IDs of the remaining tokens.
  4. Follow the previous question, is the token prune methods degrade the accuracy of the measurement of absolute distance or size. Could you provide the accuracy of diffenert category of VSI benchmark respectively?
  5. The paper states that all baselines were reproduced on Spatial-MLLM. Please provide the implementation sources, adaptation details, pruning layers, and hyperparameter configurations for each baseline to substantiate the fairness and reproducibility of the comparisons.
Reviewer_jM7zOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

WaLKPruner is a training-free selector for fused 2D–3D visual tokens. It combines scene coverage, ordered camera-path coverage, and constrained token swaps. On VSI-Bench, ScanQA, and SQA3D, it compares with pruning baselines on the same backbone at four retention ratios.

Critical strengths
  1. Same-backbone, same-budget results are strong: VSI Average reaches 50.25 at 25% retention, close to the dense model's 50.39.
  2. Figure 7 ablates scene selection, path selection, and observation order.
Critical weaknesses
  1. Figure 4 reports FLOPs but not selector runtime or end-to-end latency. At 25% retention, its reported FLOP advantage over Geo3DPruner is about 4%, so the practical speed comparison remains uncertain.
  2. Figure 7 shows a small, consistent order benefit, but does not establish whether the full path procedure beats simpler per-view or pose-stratified allocation.
Questions for authors
  1. What is the matched end-to-end latency, including selection, against Geo3DPruner?
  2. At the same budget, how does path refinement compare with simple view allocation, and how uncertain is the small Path-to-PoseSet difference?

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

Related papers

Discussion (0)

Sign in to comment.