acceptodds
Withdrawn from ICLR 2027

MAVER: Memory-Augmented View-Decoupled Evidence Routing for Video Spatial Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Video spatial reasoning requires integrating spatial cues that are distributed across time and viewpoints. Recent geometry-grounded video-language models provide increasingly rich reconstruction and camera-aware features, yet these dense representations are commonly compressed before language reasoning, which can obscure question-relevant spatial evidence. We introduce MAVER, a structured evidence interface that operates on high-resolution geometry-grounded video features before spatial pooling. MAVER first conditions current spatial features on previously observed context, then organizes them into cross-view shared and observation-specific evidence, and finally performs question-conditioned sparse routing. The selected evidence follows a dual-path injection scheme that both guides the dense visual stream before pooling and preserves selected evidence through a direct projected bypass to the language model. On VSI-Bench and VSTI-Bench, MAVER achieves average scores of 64.4 and 61.6, respectively, including a 7.1-point improvement over the controlled geometry-grounded baseline on VSI-Bench. These results suggest that structured evidence transformation before language reasoning provides an effective way to improve geometry-grounded video spatial reasoning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Spatial Reasoning, Spatial Intelligence, Vision-Language Models, Geometric Reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2sRSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to preserve and utilize geometric visual features in video spatial reasoning and proposes the MAVER framework, which consists of temporal memory, viewpoint decoupling, and problem-conditional evidence routing. Experiments achieved scores of 64.4 and 61.6 on the VSI-Bench and VSTI-Bench datasets, respectively, yielding competitive results.

Critical strengths
  1. The research question is meaningful. When geometric features are compressed and passed to language models, information related to specific spatial problems may be lost; this paper conducts a series of investigations into this issue.

  2. The experimental evaluation is detailed and comprehensive. Table 3 covers all eight combinations of the three modules, with the complete model achieving the best results; Table 4 also examines the dual-branch design, evidence selection methods, the number of evidence points, and injection paths.

Critical weaknesses
  1. Limited innovation. Temporal state updates, query-based feature aggregation, problem-conditional token selection, and residual feature injection are all relatively common designs in existing research. This paper needs to explain more clearly what new capabilities the combination of these components yields.

  2. The validation of viewpoint decoupling remains insufficient. DED uses shared queries, cross-frame aggregation, and auxiliary losses to construct two types of representations; however, these structural constraints do not directly guarantee that the representations correspond to viewpoint-invariant and viewpoint-specific information, respectively. I would like to see controlled experiments that maintain scene content while varying camera trajectories or viewpoints, as well as an analysis of how the two types of representations detect object attribute.

  3. Current ablation studies cannot fully explain the independent contributions and mechanisms of action of each module. When ER is added alone, performance drops from 57.3 to 56.6, indicating that its effectiveness clearly depends on other modules. Furthermore, regarding the claim of problem-relevant evidence selection, simply comparing learning-based selection with uniform selection is insufficient. Finally, I would like to know whether, while keeping the number of candidates, the number of parameters, and the training budget consistent, a dual-branch design with explicit perspective semantics remains superior to a standard dual-branch or single-branch design.

Questions for authors

See Weaknees.

Reviewer_6gxqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MAVER, a framework for video spatial reasoning built on VLM-3R. It introduces causal temporal memory (CTSM), dual evidence decoupling (DED), and question-guided sparse evidence routing (SER) to organize and select spatial features before pooling. The selected evidence is used to refine visual features and is also passed directly to the LLM. MAVER reports 64.4 on VSI-Bench and 61.6 on VSTI-Bench.

Critical strengths
  1. The paper addresses a relevant problem in video spatial reasoning, where fine-grained geometric evidence may be lost during feature compression.

  2. The overall architecture is clearly presented. The three modules form a coherent pipeline, and the dual-path injection provides a reasonable way to preserve selected evidence.

  3. The authors provide component-level ablations and report improvements on both benchmarks, particularly for appearance-order reasoning.

Critical weaknesses
  1. The technical novelty is limited. VLM-3R already performs geometry-visual fusion before spatial pooling, while Q-GeoMem also incorporates question-guided memory processing at this stage. MAVER mainly combines existing ideas of temporal memory, latent feature extraction, and sparse routing. Although the specific architecture differs, its contribution over these closely related approaches is not sufficiently established.

  2. The claimed view-invariant and view-specific decoupling is not convincingly validated. Shared queries and cross-frame averaging do not necessarily produce view-invariant representations. The ablations only demonstrate the benefit of using two evidence banks, without verifying whether they actually capture the intended spatial properties.

  3. The necessity of the proposed pre-pooling routing remains unclear. There is no comparison with post-pooling routing or simpler question-conditioned token compression methods. SER alone performs worse than the baseline (56.6 vs. 57.3), while adding SER to CTSM+DED improves the score by only 1.2 points. Removing the evidence bypass causes a further drop of just 0.4 points. These results provide limited support for the paper's central design choices.

  4. The improvement on VSI-Bench is largely driven by appearance-order reasoning (+37.0 points), while gains on most other categories are relatively small. MAVER also underperforms the published VLM-3R on several metric reasoning tasks. It is therefore unclear how broadly the proposed method improves spatial reasoning.

  5. The experimental comparison omits relevant recent methods, including Q-GeoMem, and SpatioLM. Moreover, the controlled baseline is not clearly matched in training compute and optimization settings.

Questions for authors
  1. What is the key method contribution beyond existing pre-pooling geometry fusion and question-guided memory approaches, particularly VLM-3R and Q-GeoMem?

  2. Can the authors provide direct evidence that the two evidence banks learn view-invariant and view-specific representations, rather than simply complementary latent features?

  3. How does MAVER compare with post-pooling routing or standard learnable-query compression under matched parameter and token budgets?

  4. Why are the improvements concentrated in appearance-order reasoning? What explains the limited gains or degradation in several metric reasoning categories?

Reviewer_Yhe7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propsed MAVER, a framework for video spatial reasoning. By design the CTSM and SER mechanisms, the framwork identify the evidence spatial token from the foundation spatial encoder and select the evidence tokens that are correlated to the question. Experiments on VSI-Bench and VSTI-Bench show state-of-the-art performance of the proposed framework.

Critical strengths
  1. The paper is well-motivated.
  2. The techniques are sound.
  3. The experiments show significant performance improvement, especially the appr. task.
Critical weaknesses
  1. On VSI-Bench, the propsed method achieves significant performance gain regarding the Appr. task, which is consistent with the claims in the paper. However, I also noticed a performance drop regarding most length/distance estimation tasks, especially on Room.S (8.6$\downarrow$), compared to the baseline VLM-3R model. Analysis should be made to explain how the proposed machenisms affect such tasks.
  2. Lacking case study and qualitative comparisons with the baseline models.
  3. Despite the ablation study conducted, it is still not unclear that the selected evidence tokens from SER are really related to the given question. It would be more persuasive if the authors can provide such evidence by providing several case study or employing related metrics.
Questions for authors

please see weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.