acceptodds
Under review as a conference paper at ICLR 2027

MindTraverse: Benchmarking Cross-Scale Spatial Composition

Abstract

Real-world spatial tasks often exhibit **cross-scale structure**: models must organize and propagate spatial context across continuous sequences of local spatial judgments and global environmental reasoning. Yet such cross-scale spatial composition has been largely overlooked by existing benchmarks, which either focus on single-step reasoning or fail to capture the cross-scale nature of real-world space in multi-step tasks. To investigate this underexplored problem, we construct **MindTraverse**. The benchmark recursively composes atomic task interfaces to create multi-step cross-scale reasoning problems validated through both geometric and visual checks. In MindTraverse, models must continuously propagate and reason over spatial context between fine-grained, object-level workspace reasoning and coarse-grained, environment-level transit reasoning. The resulting benchmark contains **6,167 multi-view questions**, covering **35 composition templates**, **six scale sequences**, and **210 simulator scenes**. Across **20 model configurations**, even the best-performing model achieves only **50.92% overall accuracy**. More importantly, endpoint accuracy does not decrease monotonically as the intended compositional dependency becomes deeper. To investigate this counterintuitive behavior, we conduct further controlled experiments and reveal three systematic deficiencies in current spatial reasoning models. First, **Structural Binding Failure**: models recover task-relevant semantic entities but struggle to organize them into coherent cross-scale structure. Second, **State–Evidence Conflict**: even when explicit spatial state is available, visual evidence is not integrated reliably and can interfere with downstream reasoning. Third, **Trajectory-Role Shortcut**: on some high-accuracy tasks, models bypass the intended cross-scale dependency and collapse onto simple trajectory roles. These results show that genuine cross-scale spatial reasoning requires not only correct final answers, but also correct spatial representation, reliable evidence integration, and faithful downstream use of that representation.

Primary area:
datasets and benchmarks
Keywords:
spatial intelligence, Multimodal learninig, VQA

Reviews

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

This work presents a spatial intelligence benchmark named MindTraverse. MindTraverse focuses on addressing the limitation of existing benchmarks that overlook the problem of cross-scale spatial composition. MindTraverse consists of approximately 6K questions spanning 35 composition templates and 210 simulator scenes. Experiments with both proprietary models and open-source models on the MindTraverse benchmark show the limitation of existing models for solving the problem of cross-scale spatial composition. In addition, the authors present in-depth analyses of counterintuitive experimental observations.

Critical strengths
  1. This paper is well-written and well-organized.
  2. Table 2 reports comprehensive experiments with both proprietary models and open-source models.
Critical weaknesses
  1. The contribution of the proposed benchmark is limited. The formalized task of cross-scale spatial composition essentially involves alternating between local spatial reasoning and global transit reasoning, which has been implicitly covered by existing benchmarks. In addition, combining multiple spatial subtasks into a comprehensive spatial task has been studied in existing benchmarks.

  2. The proposed benchmark has only approximately 6K questions, which is relatively small-scale compared with some latest spatial intelligence benchmarks.

  3. The key differences between the proposed benchmark and other recent benchmarks, such as [1,2,3], are not clearly discussed.

[1] ESI-BENCH: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop. Hong et al. 2026.

[2] How Far are VLMs from Visual Spatial Intelligence? A Benchmark-Driven Perspective. Yu et al. 2025.

[3] Spatial4D-Bench: A Versatile 4D Spatial Intelligence Benchmark. Wang et al. 2026.

Questions for authors
  1. For cross-scale spatial composition, the proposed benchmark primarily covers W, T, W-T, T-W, W-T-W, and T-W-T. Why are longer and more complex sequences not included in the proposed benchmark?
  2. Does the proposed benchmark only focus on static scenarios? Why is cross-scale 4D spatial reasoning not considered in the proposed benchmark?
11 comments

Q1

Thank you for suggesting that we explain the choice of sequence lengths. We agree that we should clarify the reasons for the current sequence setting. The six sequences include two atomic tasks, the two cross-scale transition directions (WT and TW), and three-stage tasks that return to the original scale (WTW and TWT). They allow us to examine cross-scale handoffs and subsequent state use separately and associate errors with specific stages. We do not use longer sequences for the following reasons:

(1) Questions with longer sequences can be viewed as compositions of the current six sequences. The current six sequences already cover all possible input conditions and evaluation forms for W and T atomic questions; longer sequences further compose these existing questions.

(2) Longer sequences do not introduce additional diagnostic evidence. In finite indoor layouts, simply extending a chain can introduce repeated entry into and exit from the same room without additional information. We prioritize explicit information dependencies at every transition, facilitating diagnosis of cross-scale spatial reasoning.

(3) The current benchmark already fully exposes models’ bottlenecks in cross-scale spatial reasoning. Many models perform near chance on three-stage questions. Further increasing difficulty by extending the reasoning sequence would not better reveal their current bottlenecks in cross-scale spatial composition.

(4) The choice of current sequence formats is not a technical limitation. Our benchmark can generate questions of arbitrary length. Through recursive composition and scene grounding, we can generate sequences and questions of arbitrary length; the current formats therefore do not reflect a technical limitation.

Q2

Thank you for suggesting that we clarify the relationship to 4D spatial reasoning. We agree that your concern about possible 4D spatial reasoning is reasonable. Accordingly, we have added a comparison with 4D spatial reasoning to Related Work in the main text.

Our benchmark belongs to the research direction of multi-view spatial intelligence, represented by works such as MindCube [4] and MMSI-Bench [5]. This direction primarily examines whether models can connect viewpoints through shared entities and spatial cues. 4D spatial intelligence instead uses RGB video as input and examines spatial intelligence under temporal evolution.

We choose static multi-view images for the following main reasons:

  1. We want to focus on spatial relations and reference-frame linkage.

    Cross-scale spatial reasoning examines whether models can link small-scale and large-scale reference frames and flexibly switch between them. Multi-view spatial intelligence closely matches this central question and offers a setting with fewer additional factors. If cross-scale evaluation were built on 4D spatial intelligence, models would also need to handle temporal order, object changes, and related factors, which would interfere with diagnosing cross-scale spatial modeling.

    In our concrete task rules, all long-chain tasks require room transitions. We do not provide complete observation sequences for every room; instead, we examine whether models can build an overall spatial model from sparse views through shared entities. Video provides more complete observations of entities, reducing the demand on this ability while potentially introducing additional factors.

  2. Multi-view inputs make reference-frame specification explicit.

    Our benchmark defines the reference frame either by facing a specified anchor object or by using an explicitly numbered input image. Providing selected views directly allows us to control the available observations and include the exact image referenced by a question, avoiding the risk of losing that view during video-frame sampling.

Nevertheless, 4D spatial intelligence remains an important direction, as you pointed out. We have therefore added a comparison with 4D spatial reasoning to Related Work.

W1

Thank you for pointing out that existing compositional tasks need to be compared more explicitly. We have already addressed this comparison in Related Work: MMSI-Bench [5] includes Multi-Step Reasoning tasks that chain spatial judgments, and SpaCE-10 combines ten atomic capabilities into eight QA types.

MindTraverse differs by explicitly binding typed intermediate results as conditions for the next-scale operation. For example, a direction from the source workspace determines the subsequent entrance, and the transit destination then determines the terminal workspace. The generation program records these dependencies, and supplied-state and paired interventions examine whether models use them.

MindTraverse also supports recursive composition and can generate questions with sequences of arbitrary length, a capability not provided by the above benchmarks.

W2

Thank you for your concern about data scale and coverage. MindTraverse contains 6,167 evaluation questions covering 210 scene instances, 35 templates, and six scale sequences. Among fixed multi-view spatial evaluations, MMSI-Bench contains 1,000 questions and SpaCE-10 [6] contains 5k+. Our question count is therefore comparable to these related benchmarks, and we believe the current scale is sufficient to diagnose cross-scale spatial composition.

The current scale also leaves room for expansion. The 210 scene instances come from 185 distinct source houses within the 10,000-house ProcTHOR-10K training set. Through additional compatible scenes, entity combinations, and initial states, recursive composition and scene grounding can automatically generate further questions that satisfy the validation requirements. Our generation method is therefore scalable and, if necessary, can produce far more than 6,167 questions.

W3

Thank you for recommending these three related works. We have cited them in Related Work and explained their evaluation focus and relationship to this work. We have also added comparisons with these papers to the main text:

WorkInput formatMain evaluation focus
ESI-Bench [1]Interactive RGB observationsEvaluating the perception–action loop through active observation and environmental interaction
SI-Bench [2]Single images, multiple images, and videosGeneral spatial capabilities, including perception, understanding, and planning
Spatial4D-Bench [3]VideoSpatial understanding and reasoning involving temporal changes
MindTraverse (Ours)Multiple imagesExplicit cross-scale result dependencies in multi-view inputs

MindTraverse focuses on explicit cross-scale result dependencies in fixed multi-view inputs and uses controlled interventions to examine whether models use those dependencies.

[1] ESI-BENCH: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop. Hong et al. 2026.

[2] How Far are VLMs from Visual Spatial Intelligence? A Benchmark-Driven Perspective. Yu et al. 2025.

[3] Spatial4D-Bench: A Versatile 4D Spatial Intelligence Benchmark. Wang et al. 2026.

[4] Spatial Mental Modeling from Limited Views. Wang et al. 2026

[5] Mmsi-bench: A benchmark for multi-image spatial intelligence. Yang et al. 2026

[6] SpaCE-10: A Comprehensive Benchmark for Multimodal Large Language Models in Compositional Spatial Intelligence. Gong et al. 2025

Thanks for the responses.

First, using intermediate results as the conditions for subsequent reasoning is not a NEW idea. It remains unclear how the proposed composition FUNDAMENTALLY differs from existing works.

Second, given that MindTraverse is comparable in scale to existing benchmarks and similar tasks could potentially be constructed using existing benchmarks or simulators, what justifies introducing MindTraverse as a new benchmark? What unique value and impact would it bring to the community?

Reviewer_9NfsOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MindTraverse, a benchmark for evaluating cross-scale spatial composition in multimodal models. The benchmark requires models to continuously transfer spatial information between workspace-level reasoning and transit-level reasoning. It contains 6,167 questions covering 35 composition templates and six scale sequences, and the authors evaluate 20 model configurations. The experiments show that current models still struggle with this type of spatial reasoning, and the paper further analyzes three failure modes, including structural binding failure, state-evidence conflict, and trajectory-role shortcut.

Critical strengths
  1. This paper is well written and easy to follow;
  2. The paper studies an interesting problem which is not well covered by existing spatial reasoning benchmarks.
  3. The benchmark construction is careful, with both geometric validation and visual answerability checking.
  4. The diagnostic experiments are interesting and provide more insights than only reporting final accuracy.
Critical weaknesses
  1. The benchmark is still based on simulated household scenes with fixed observations, so the generalization to real-world or embodied settings is unclear.
  2. Many detailed diagnostic experiments are mainly conducted on Qwen3.7-Flash and selected subsets, so I am not sure whether these failure modes are general across different models.
  3. Different models use different prompts, reasoning settings, and inference budgets, which makes the comparison not fully fair.
  4. Some task structures seem to have exploitable shortcuts or restricted answer spaces, so the final accuracy may not always reflect the intended cross-scale spatial reasoning ability.
Questions for authors

Could the authors provide similar diagnostic experiments on more model families, instead of mainly focusing on Qwen3.7-Flash?

2 comments

Q1 & W2

Thank you for the suggestion. We extended the map and state diagnostics in Table 3 to Gemini 3.7 Flash and Qwen3.8 Flash. The two additional models use the same 300-question subset, with 50 questions per scale sequence; Qwen3.7’s results are also drawn from these same 300 questions. The table reports accuracy (%). Mapping RGB and Reasoning RGB indicate whether images are supplied during map generation and downstream reasoning, respectively.

PredMapGT MapPoseMapping RGBReasoning RGBQwen3.7 Flash (300 questions)Gemini 3.7 Flash (300 questions)Qwen3.8 Flash (300 questions)
Representation controls
✓––✓–44.357.045.0
✓–✓✓–35.751.742.0
–✓✓––53.386.763.3
State–evidence integration
–✓✓––53.386.763.3
–✓✓–✓45.085.047.3

All three models improve when the pose-conditioned predicted map is replaced with the GT map. Holding GT Map + Pose fixed, adding RGB reduces accuracy by 8.3, 1.7, and 16.0 percentage points for Qwen3.7, Gemini, and Qwen3.8, respectively. Gemini performs better and is less affected by RGB, consistent with its leading performance in the main table. These results extend the evidence that explicit spatial state helps reasoning to additional models and show model differences in performance when state and visual evidence are combined.

The experimental setup and results for the same 300 questions have been added to Appendix D.9, Table A21; the shared mapping and reasoning protocol is provided in Appendix B.2 and H.2.

W1

Thank you for your concern about generalization to real-world scenes. The simulator provides accurate ground truth for cameras, objects, and spatial geometry, enabling us to instantiate cross-scale dependencies, validate answers, and perform controlled interventions. In future dataset development, we will consider extending to real-world scenes with reliable 3D annotations.

W3

Thank you for pointing out the differences in model configurations. The main experiment provides each model with the same forms of public input: ordered RGB images, the question, and options. The prompts and reasoning settings are given in Appendix B. Gemini and GLM cannot disable thinking mode, so we use the lowest reasoning setting allowed by their providers; thinking is disabled for the other models. Reasoning budgets therefore differ, as noted in the Table 2 caption. This does not affect our diagnosis of models’ cross-scale spatial composition.

The map diagnostics use task-specific prompts and shared reasoning rules within each model; the specific controlled input changes are described in Appendix B.2, D.9, and H.2.

W4

Thank you for pointing out how shortcuts affect the interpretation of results. We should examine shortcut reliance alongside accuracy. Our cross-model results show that higher benchmark accuracy co-occurs with lower shortcut following. Gemini achieves higher accuracy in the main evaluation while following the trajectory-role shortcut less often. In the identity swap on 77 TWT questions, unvisited-role following is 98.7% for Qwen3.7 and 42.9% for Gemini. We conclude that evaluating model capability requires considering both shortcut use and accuracy. However, models with higher accuracy also tend to have lower shortcut reliance, so accuracy can partly reflect cross-scale capability.

After carefully reviewing the authors' rebuttal and the comments from other reviewers, I have decided to maintain my original score.

Reviewer_7wbPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper tests whether models can link local spatial judgements with room level decisions. It introduces 6k+ questions and evaluates 20 model configurations, using human evaluation and controlled tests to check the tasks and examine errors.

Critical strengths
  1. The shortcut interventions are convincing and help explain some models’ successes and failures.
  2. The construction checks and human baseline support the data quality and make the benchmark more credible.
Critical weaknesses
  1. Some questions may be answerable from the final images and object names alone, so the earlier steps may not be necessary.
  2. Models often struggle with basic spatial judgements, making it hard to identify the extra difficulty introduced by cross-scale composition.
Questions for authors
  1. Clarification of the design: What existing checks rule out answering from the final images alone?
  2. Clarification of the results: Is there further evidence of composition failures when all required basic spatial judgements are correct?
1 comment

Q1 & W1

Thank you for raising this question. The construction program binds preceding results to the selection of subsequent entrances, destination rooms, or reference frames, while geometric and visual checks validate the answers and the required evidence. To further examine the actual contribution of preceding images to model answers, we added a terminal-image ablation experiment.

We selected 70 questions from different scenes for each of WT, TW, WTW, and TWT, for a total of 280 questions, and compared full-image input (Full) with retaining only the last three, two, or one image. Every original input contains at least four images. The question text, options, and original indices of retained images remain unchanged across conditions.

ModelFullLast 3 imagesLast 2 imagesLast imageMean relative drop vs. Full
DeepSeek V4.1 Flash-101743.57%35.71%36.79%38.21%15.30%
Qwen3.7 Flash45.36%34.64%36.07%38.57%19.69%
Qwen3.8 Flash37.14%32.50%31.07%30.71%15.38%
Overall (AVG)42.02%34.29%34.64%35.83%16.90%

For all three models, overall accuracy under each terminal-image condition is lower than under Full, indicating that preceding images contribute to solving these tasks.

Our specific constraints on image organization are as follows:

  1. WT tasks require local information from the workspace in the first photograph to determine the next action.
  2. In TW and WTW tasks, the last three photographs provide only sparse observations of the corresponding workspace and do not establish whether it is the workspace specified in the question.
  3. In TWT tasks, the last three images comprise a route-connection view and two photographs of landmark objects associated with the candidate rooms.

Construction and evidence checks are described in Appendix A.3–A.6. The new terminal-image ablation has been added to Appendix D.7, Table A19.

Q2 & W2

Thank you for highlighting this key distinction. We added a matched comparison between independent subtasks and the full task to test whether models can correctly answer all required spatial judgments yet still fail on the original multi-stage task.

We split the same 100 WTW questions into three independent subtasks: W1 judges the source workspace relation; T receives the correct W1 result and selects the target room; W2 receives the correct target room and an additional rendered photograph, and directly judges the terminal relation from the photograph’s camera viewpoint. The horizontal orientation of the W2 photograph matches the original question’s requirement, so the model need not imagine facing the anchor. Full WTW retains the original question and images and still requires the instructed hypothetical arrival and orientation.

ModelQuestionsAll three independent subtasks correct on responseOriginal WTW wrong on response among these
Qwen3.7-Flash10095
Qwen3.8-Flash100118
DeepSeek V4.1 Flash-101710096

All three models exhibit cases where every independent subtask is correct but the full task is still wrong. These counterexamples show that completing the required spatial judgments independently does not guarantee that models can correctly compose and apply them under the original multi-stage instruction.

These failures arise because cross-scale transitions also test whether models can model what would be seen from a viewpoint not shown in the images. In our task setting, when a W task follows a T task, the model cannot reason directly from the given images: it must locate the anchor object and model the scene as viewed while facing that anchor before reasoning. Our atomic W2 task skips this imagined-viewpoint step. This leads to the models’ failures and confirms that even when models can complete all local reasoning, they can still fail to solve the full reasoning chain.

The experimental setup and results have been added to Appendix D.8, Table A20.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Responses to Reviewers

Summary of Our Responses to Reviewers

We thank the three reviewers for their careful reviews and positive comments on the paper's presentation, dataset construction, and diagnostic analyses. We have added experiments and technical clarifications addressing their main concerns, summarized below.

Response to Reviewer 7wbP: We added last-view ablations on 280 questions and independent subtask controls on 100 questions across three models. These show the contribution of earlier observations and cases where all subtasks are correct but the full task is wrong. We also clarify that the independent W₂ probe receives an additional image aligned with the target facing direction, while the full task must infer that view from the original observations.

Response to Reviewer 9Nfs: We extended map diagnostics to Gemini and Qwen3.8 on the same 300 questions to compare the effects of explicit state and RGB, and included DeepSeek in other controls. We also clarified model configurations, reasoning budgets, and scope, and used Text-only and role interventions to explain shortcut use and differences across models.

Response to Reviewer NEyL: We added related-work comparisons and technical explanations of task generation and dual geometric and visual validation, clarifying that MindTraverse goes beyond task composition to assess models’ ability to jointly model spatial structure across scales. We also added 200 four- and five-stage questions, evaluations of four model configurations, and worked examples, and used task structure, shortcut diagnostics, and the human baseline to address sequence performance trends and benchmark reliability.

The new experiments are detailed in Appendix Sections D.7–D.10 of the revised manuscript. We again thank the reviewers and the Area Chair for their time and suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.