Spatial reasoning, the ability to understand spatial relations, causality, and dynamic evolution, is central to human intelligence and essential for real-world applications such as autonomous driving and robotics. Existing studies, however, primarily assess models on visible spatio-temporal understanding, overlooking their ability to infer unseen past or future spatial states. In this work, we introduce Spatial Causal Prediction (SCP), a new task paradigm that challenges models to reason beyond observation and predict spatial causal outcomes. We further construct SCP-Bench, a benchmark comprising 2,500 QA pairs across 1,181 videos spanning diverse viewpoints, scenes, and causal directions, to support systematic evaluation. Through comprehensive experiments on {23} state-of-the-art models, we reveal substantial gaps between human and model performance, limited temporal extrapolation, and weak causal grounding. We further analyze key factors influencing performance and propose perception-enhancement and reasoning-guided strategies toward advancing spatial causal intelligence. The project page is https://guangstrip.github.io/SCP-Bench.
Figures & tables
Figure 1: Existing benchmarks primarily assess known static or known dynamic reasoning based on fully observable scenes. A more challenging dynamic–unseen setting is to evaluate models’ ability to predict spatial outcomes from partial observations.
Figure 2: Overview of SCP-Bench. Left : Representative examples illustrating the eight task categories. Right : Data distribution across scene categories and task types. The benchmark comprises 2,500 QA pairs over 1,181 video clips.
Figure 3: Overview of the SCP-Bench construction pipeline. The process comprises five stages: (1) collection of diverse video sources, (2) clip selection with spatially dynamic segments, (3) generation of candidate QA pairs, (4) QA filtering and cutpoint identification, and (5) dataset validation and refinement.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
Human Performance
89.61
97.60
81.20
92.26
85.70
86.70
97.62
91.61
84.17
∙ Closed Models
GPT-5
66.24
79.04
58.12
59.06
64.07
70.48
95.24
77.42
65.11
Gemini-2.5-Pro
55.84
69.28
54.87
52.76
46.20
63.47
88.10
67.10
62.41
Gemini-2.5-Flash
52.10
59.28
52.14
51.74
43.14
57.75
88.10
66.45
55.60
Claude-Sonnet-4.5
56.14
68.86
52.14
57.43
45.65
60.90
80.95
68.39
63.90
Table 1: Evaluation on the SCP-Bench. “Avg.” indicates the overall average accuracy. For each category, the best-performing closed model and open-source model in average score are both indicated in deep blue , and best performance on each task is b o x e d .
Figure 4: Results across perspectives, view directions, and scenes.
Figure 5: Temporal extrapolation horizon analysis. Samples are grouped by the time gap between the cutpoint and future event: short (0–2s), mid (2–5s), and long ( > 5s).
Setting
Qwen3-VL
InternVL3.5
MiniCPM-V-4.5
Base
47.52
50.52
43.80
Gold Video
54.96 ↑7.44
58.64 ↑8.12
43.72 ↓0.08
Caption w/o Video
46.76 ↓0.76
41.24 ↓9.28
45.76 ↑1.96
Table 2: Comparison between perception and reasoning. “Base” is the standard evaluation performance; “Gold Video” evaluates perceptual understanding using unseen part; “Caption w/o Video” tests reasoning based on captions alone. Red and green indicate performance gains and drops, respectively.
Figure 6: Visible Range Comparison. Only cutpoint uses the cutpoint frame as input; Full video provides the entire clip; Ground truth includes only the unseen clips adjacent to the cutpoint.
Figure 7: Performance gap comparison. Accuracy improvements over the chance level for vision-enabled (w/ video) and vision-disabled (w/o video) settings across task categories.
Figure 8: Causal consistency evaluation. Each option shows the model’s reasoning rationality, confidence score, and explanation, assessing whether MLLMs can infer spatial order consistent with physical and causal constraints.
Setting
Qwen3-VL
InternVL3.5
MiniCPM-V-4.5
LLaVA-OV-1.5
Base
47.52
50.52
43.80
45.52
w/ Flip
46.76 ↓0.76
48.56 ↓1.96
42.24 ↓1.56
45.32 ↓0.2
w/ CoT
51.52 ↑4.0
50.36 ↓0.16
44.88 ↑1.08
42.88 ↓2.64
Table 3: Comparison with flipped video input and CoT reasoning. “w/ Flip” reverses the input video temporally; “w/ CoT” applies step-by-step reasoning.
Figure 9: Performance comparison on different model sizes.
Figure 10: Performance with vs. without self-think reasoning.
Figure 11: Comparison of perception enhancement strategies. PureText uses only the question; Caption w/ V combines dense video captions with the video and question; Caption w/o V uses only captions with the question; SpatialGraph introduces spatial-interaction graph; Original is the baseline.
Figure 12: Unseen spatial causal scaffolds for enhanced reasoning. Text provides textual descriptions of future spatial states; Image employs generated future spatial images; Video uses generated future causal videos.
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
Category
Video Clip
QA Instance
∙ Scene Category
Sports Related
467
964
Life Records
289
691
Driving-view Related
122
308
Animal Related
85
143
Artistic Performances
64
95
Appendix
Table 4: Distribution of benchmarks over the scene category, perspective, Causal direction, and question type. These attributes are combined to form the final dataset, which consists of 1,181 video clips and 2,500 QA pairs.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Gold Video
Qwen3-VL-8B
54.96 ↑7.44
62.28
58.12
59.47
48.17
54.26
83.33
57.42
57.91
InternVL3.5-8B
58.64 ↑8.12
68.86
57.26
64.97
51.37
60.37
80.95
60.00
58.27
MiniCPM-V-4.5
43.72 ↓0.08
48.50
47.01
44.40
35.47
50.80
78.57
54.19
43.53
∙ Dense Caption Only
Qwen3-VL-8B
46.76 ↓0.76
51.50
37.61
53.97
40.16
48.94
80.95
60.00
40.29
Appendix
Table 5: Evaluation results for the “Perception vs. Reasoning Decomposition” experiment. “Avg.” indicates the overall average accuracy. Red and green indicate performance gains and drops relative to the Baseline setting, respectively.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Cutpoint Image
Qwen3-VL-8B
49.52 ↑2.00
49.70
52.14
50.92
44.62
50.27
83.33
55.48
51.80
InternVL3.5-8B
51.92 ↑1.40
67.07
52.99
52.14
44.16
57.18
76.19
61.29
50.36
MiniCPM-V-4.5
43.80 0.00
52.69
55.56
44.60
35.35
50.80
76.19
50.97
40.29
Spatial-MLLM
42.04 ↑2.28
53.46
32.73
38.90
37.30
53.55
71.43
55.69
33.33
Appendix
Table 6: Evaluation results for the “Single-Frame vs. Multi-Frame Reasoning” experiment.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Reverse Video
Qwen3-VL-8B
46.76 ↓0.76
52.69
52.14
46.84
43.94
46.54
90.48
48.39
42.45
InternVL3.5-8B
48.56 ↓1.96
56.89
56.41
49.49
42.33
53.19
64.29
59.35
43.53
MiniCPM-V-4.5
42.24 ↓1.56
49.10
45.30
41.96
35.13
48.67
73.81
50.97
41.37
LLaVA-OneVision-1.5-8B
45.32 ↓0.20
52.69
47.01
45.42
39.47
50.00
80.95
49.03
44.60
Appendix
Table 7: Evaluation results for the “Vision Causal Perception” experiment.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Baseline
Qwen3-VL-8B
47.52
54.49
51.28
49.29
42.33
49.47
90.48
46.45
46.40
InternVL3.5-8B
50.52
59.88
54.70
54.79
43.82
54.52
61.90
58.71
44.96
MiniCPM-V-4.5
43.80
53.29
49.57
43.99
36.04
49.20
76.19
52.26
42.81
Spatial-MLLM
39.76
45.51
28.21
33.81
38.33
49.73
66.67
50.97
32.37
∙ Gold Video
Appendix
Table 8: Evaluation results for the “Visible Range Comparison” experiment.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Pure Text
Qwen3-VL-8B
42.00 ↓5.52
49.10
35.04
47.25
34.67
46.28
69.05
40.65
45.32
InternVL3.5-8B
44.72 ↓5.80
50.90
42.74
48.88
36.61
52.66
73.81
51.61
41.01
MiniCPM-V-4.5
43.32 ↓0.48
53.29
44.44
46.03
35.01
51.60
69.05
52.90
37.77
LLaVA-OneVision-1.5-8B
40.20 ↓5.32
43.11
41.03
42.16
33.41
47.87
73.81
38.71
41.37
∙ Dense Caption Only
Appendix
Table 9: Evaluation results for the “Is Visual Information Necessary?” experiment.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Qwen3-VL Series
Qwen3-VL-2B
43.04
41.92
42.74
45.01
40.85
44.41
59.52
47.10
40.65
Qwen3-VL-4B
50.24
58.68
47.86
52.55
44.51
55.05
90.48
60.00
42.09
Qwen3-VL-8B
47.52
54.49
51.28
49.29
42.33
49.47
90.48
46.45
46.40
Qwen3-VL-30B-A3B
54.16
65.27
52.14
54.79
46.22
56.65
85.71
66.45
57.19
Qwen3-VL-32B
56.84
59.88
51.28
58.66
52.63
57.98
90.48
67.10
55.04
Appendix
Table 10: Evaluation results for the “Model Scale-up Effect Analysis” experiment.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Vanilla CoT
Qwen3-VL-8B
51.52 ↑4.00
63.47
50.43
51.53
44.39
54.52
90.48
61.29
51.80
InternVL3.5-8B
50.36 ↓0.16
59.28
51.28
49.29
43.02
56.65
61.90
65.16
51.08
MiniCPM-V-4.5
44.88 ↑1.08
55.09
40.17
48.68
35.47
53.19
73.81
59.35
39.93
LLaVA-OneVision-1.5-8B
42.88 ↓2.64
48.50
39.32
46.44
33.30
53.99
69.05
54.19
39.57
Appendix
Table 11: Evaluation results for the “CoT Reasoning” experiment.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Self-Thinking
Qwen3-VL-8B
47.44 ↓0.08
55.69
45.30
50.10
38.79
51.86
83.33
58.71
48.20
InternVL3.5-8B
47.84 ↓2.68
59.28
51.28
51.12
39.36
53.46
66.67
54.19
46.40
MiniCPM-V-4.5
42.72 ↓1.08
49.10
48.72
40.94
35.35
50.53
69.05
50.32
43.88
Appendix
Table 12: Evaluation results for the “Self-Think Reasoning” experiment.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Dense Caption Auxiliary
Qwen3-VL-8B
47.72 ↑0.20
54.49
48.72
52.75
41.76
46.01
88.10
53.55
46.04
InternVL3.5-8B
50.56 ↑0.04
60.48
53.85
54.18
43.02
56.12
64.29
60.65
45.32
MiniCPM-V-4.5
44.28 ↑0.48
56.29
45.30
51.73
33.64
46.28
85.71
54.84
42.09
LLaVA-OneVision-1.5-8B
45.80 ↑0.28
53.89
46.15
49.08
39.13
47.61
83.33
54.84
42.81
∙ Spatial Interaction Graph
Appendix
Table 13: Evaluation results for the “Perception Enhancement Strategy” experiment.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Text Scaffold
Qwen3-VL-8B
61.23 ↑13.71
64.29
56.67
67.44
54.48
64.08
91.89
64.06
56.25
InternVL3.5-8B
61.72 ↑11.20
64.29
57.78
67.44
53.25
65.49
91.89
67.97
59.13
MiniCPM-V-4.5
60.15 ↑16.35
58.16
52.22
66.28
51.32
64.08
97.30
67.19
59.62
LLaVA-OneVision-1.5-8B
59.77 ↑14.52
55.10
53.33
66.05
50.09
62.68
89.19
71.09
62.02
∙ Image Scaffold
Appendix
Table 14: Evaluation results for the “Causal Prediction Enhancement Strategy” experiment.
Benchmark
QA pairs
Modality
Dynamic/Static
View Type
Perspective
Causal Reasoning
Seen/Unseen
Single
Multi
Ego
Exo
3DSRBench [ 33 ]
3,772
Image
Static
✓
✗
✗
✓
✗
Seen
InternSpatial-Bench [ 15 ]
6,008
Image
Static
✓
✗
✗
✓
✗
Seen
OmniSpatial [ 27 ]
∼ 8,400
Image
Static
✓
✗
✓
✓
✗
Seen
Spatial457 [ 46 ]
23,752
Image
Static
✓
✗
Undeclared
✗
Seen
All-Angles Bench [ 57 ]
∼ 2,100
Image
Static
✗
✓
Undeclared
✗
Seen
Appendix
Table 15: Comparison with existing benchmarks. Modality denotes the input type. Dynamic/Static indicates whether the scene content itself undergoes temporal changes, rather than mere camera movement. View Type specifies whether the questions involve single or multiple viewpoints. Perspective denotes whether the benchmark includes explicitly designed ego (first-person) and exo (third-person) perspective settings. Causal Reasoning indicates whether the benchmark requires inferring outcomes or states driven by causal dependencies. Seen/Unseen reflects whether the queried information is directly observable within the given visual content.
Model
Avg. ↓
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Misclassifying the Correct Answer as Least Plausible
Qwen3-VL-8B
29.76
34.73
4.27
26.27
27.35
42.29
21.43
29.68
35.61
InternVL3.5-8B
30.96
32.34
8.55
26.27
32.49
38.83
33.33
28.39
33.45
MiniCPM-V-4.5
34.36
37.72
15.38
28.11
32.84
45.48
35.71
39.35
38.13
LLaVA-OneVision-1.5-8B
31.64
36.53
2.56
32.38
26.89
46.28
33.33
31.61
34.53
Appendix
Table 16: Results of the “Extended Causal Consistency” experiment, where models are evaluated by how often they misclassify the correct answer as the least plausible option. Lower values indicate stronger causal consistency. “Avg.” denotes the overall average error rate.
Model
Avg.
Appearance Order
Counting
Planning
Relation
Relative Distance
Relative Size
Relative Speed
Spatial State
∙ Vanilla CoT
Qwen3-VL-8B
51.52 ↑4.00
63.47
50.43
51.53
44.39
54.52
90.48
61.29
51.80
InternVL3.5-8B
50.36 ↓0.16
59.28
51.28
49.29
43.02
56.65
61.90
65.16
51.08
MiniCPM-V-4.5
44.88 ↑1.08
55.09
40.17
48.68
35.47
53.19
73.81
59.35
39.93
LLaVA-OneVision-1.5-8B
42.88 ↓2.64
48.50
39.32
46.44
33.30
53.99
69.05
54.19
39.57
∙ Physics-aware CoT
Appendix
Table 17: Evaluation results for the “Physical Commonsense Probing” experiment. “Avg.” indicates the overall average accuracy.
Figure 13: Example of causal prediction enhancement. Here we present a comparative analysis of a case study, showing the actual subsequent video and the causal prediction results for text, image, and video modalities.
Figure 14: Analysis sample for extended causal consistency experiment. The models are required to identify the least plausible option, while InternVL3.5-8B does not independently assess the plausibility of each choice but instead derives the least plausible option by reasoning backward from its own incorrect answer, revealing a lack of internal causal consistency.
Figure 15: Visualization of the physical commonsense probing experiment. To answer each question, models must first infer the underlying physical principles depicted in the video. Although current models can often recognize the appropriate physical laws, they still struggle to apply them effectively when generating correct predictions.
Figure 16: Cases of dynamic integration failure. The models fail to perceive dynamic changes and remain confined to static observations.
Figure 17: Cases of dynamic integration failure and prior-driven hallucination. The models struggle to perceive and infer the correct global spatial dynamics and often overlook critical scene details during reasoning due to prior-driven biases.
Figure 18: Cases of causal reasoning failure and cross-modal attribution bias. The models struggle to perform causal reasoning and instead produce unsupported conclusions, leading to erroneous predictions. Moreover, they may overlook information from other modalities and fall into single-modality reasoning.
Figure 19: Manual filtering tool demonstration. Annotators use it to filter appropriate QA candidates and determine their cutpoints.
Figure 20: Validation tool demonstration. Annotators use it to validate each item in strict accordance with the SCP task specification.
Figure 21: Repairing tool demonstration. Annotators use it to repair and optimize any attribute of the item as needed.
Predicting future spatial states supports collision avoidance and timely decision-making in dynamic environments. However, existing vision-language models (VLMs) and benchmarks for spatial reasoning primarily focus on observed scenes, leaving predictive spatial reasoning beyond the observed interval underexplored. To this end, we introduce SpatialMind, a metric-scale VLM for spatial reasoning and future prediction. Its metric depth adapter anchors spatial reasoning to real-world scale, while its progressive state chain establishes current spatial states and observed dynamics as the foundation for future prediction. Given a video prefix, SpatialMind predicts distances, motion directions, and spatial relations in both observed and unseen future frames. For training and evaluation, we build a scalable data engine that grounds entity descriptions in metric geometry to generate question-answer pairs and state supervision. Using this engine, we construct the SpatialMind-30K dataset and the SpatialMind-2K benchmark, both covering driving and everyday egocentric scenes. The benchmark spans eight tasks across three levels: current-state understanding, observed-dynamics understanding, and future prediction. Experiments show that SpatialMind substantially outperforms both general and spatially specialized models on our benchmark while achieving competitive zero-shot performance on VSI-Bench, OSI-Bench, and VLM4D.
Feiran Wang, Xiaoqi Wang, Ziwei Li +3
University of Illinois at Chicago · Bosch Research North America · Bosch Center for Artificial Intelligence (BCAI)
Existing spatial reasoning benchmarks mainly test spatial perception: reading off relations already visible in the input. Yet real-world spatial intelligence demands predictive spatial reasoning: constructing a scene from observations, anticipating how an intervention changes it, and reasoning about the unseen outcome. We introduce SpaceCast-Bench, the first benchmark to directly and diagnostically evaluate this capability. Built around an observe-transform-infer framework, its 3,862 questions from 182 real-world scenes span 16 task types at three levels: static perception, local prediction, and global prediction, progressively requiring scene understanding, spatial state updating, and relational inference over unobserved outcomes. Evaluating 21 models exposes a stark gap: the strongest model reaches only 58.0% against 87.2% human performance, while spatially specialized models remain near random chance. Controlled analyses further reveal that bridge views are critical for integrating distributed observations, and that explicit 3D evidence benefits models more reliably than generated outcome images or videos. Fine-tuning on our programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7% with macro-average gains across six out-of-domain benchmarks.
Cause-and-effect reasoning in video is a significant challenge for Vision-Language Models (VLMs), as it requires going beyond surface-level perception to a deeper understanding of causal mechanisms. However, existing benchmarks rarely provide the fine-grained, grounded evidence needed to rigorously evaluate this capability. To address this gap, we introduce CaST-Bench, a benchmark for Causal Chain-Grounded Spatio-Temporal Video Reasoning. CaST-Bench presents complex causal questions that require models to identify and localize a chain of multiple spatio-temporal evidences. Through a human-AI collaborative pipeline, we construct a high-quality dataset of 2,066 questions over 1,015 videos, with causal chains annotated by temporal segments and bounding-box tracks. Furthermore, we design a comprehensive evaluation suite with novel metrics that assess not only answer correctness but also the capability for visual evidence grounded reasoning. This grounding is crucial for improving accuracy by mitigating spurious correlations and for enhancing user trust by making models more transparent. Our experiments show that current VLMs struggle with causal questions, largely due to their limited ability to construct precise and grounded causal chains. This highlights an important direction for improving future VLMs. Homepage: https://woven-by-toyota.github.io/CaST-Bench.