Seek-and-View Reasoning for Multi-View Spatial Understanding

Qixiang Chen; Cheng Zhang; Fucai Ke; Chi-Wing Fu; Jianfei Cai; Jingwen Ye

cs.CVOct 8, 2026

Seek-and-View Reasoning for Multi-View Spatial Understanding

Authors: Qixiang Chen, Cheng Zhang, Fucai Ke, Chi-Wing Fu, Jianfei Cai, Jingwen Ye

Organizations: Monash University · The Chinese University of Hong Kong

Abstract

Existing approaches to multi-view spatial reasoning operate largely on sparse input views. Vision-language models (VLMs) are thus restricted to understand a scene and infer spatial relations within these fixed views, leading to fragile cross-view alignment and geometry-to-language bottleneck. To address these issues, we formulate a novel Seek-and-View reasoning approach to find implicit cross-view spatial evidence by locating a question-relevant view to support the spatial reasoning. To realize this approach, we propose Vantage, a training-free model-agnostic reasoning framework that pairs a VLM with a 3D foundation model: a viewpoint-grounded reasoning stage for question analysis and view planning, followed by a geometry-grounded evidence augmentation stage to effectively synthesize and incorporate visual evidence into the final reasoning. Comprehensive experiments on six VLMs demonstrate consistent improvements on five benchmarks without fine-tuning. Overall, by revealing spatial evidence through view-grounded reasoning, Vantage can largely reduce reliance on language-based cross-view alignment and improve multi-view spatial understanding. Our code is available at https://github.com/q1xiangchen/Vantage.

Figures & tables

Figure 1: (a) Existing approaches reason over fixed views, suffering from fragile cross-view alignment and the geometry-to-language bottleneck. (b) Our novel Seek-and-View reasoning paradigm seeks a question-relevant view (see bottom right) to make the spatial evidence directly observable.

Figure 2: Overview of Vantage pipeline. Stage 1: Viewpoint-grounded reasoning employs an VLM (in blue) to (a) produce an analysis on the input question (top-right), select reference view $v_{r}$ , identify queried spatial relation, then (b) plan a question-relevant camera action $\Delta v$ on reference view $v_{r}$ and generate reasoning guidance for supporting view interpretation. Stage 2: (c) Geometry-grounded evidence augmentation synthesizes the planned view $v_{s}$ from the 3DFM reconstruction (in green) and provides it together with the original images, viewpoint analysis $c_{\mathrm{ana}}$ , and reasoning guidance $c_{\mathrm{guide}}$ to support (d) the final reasoning.

Model	MindCube-tiny				MMSI-Bench			BLINK	OmniSpatial	SPINBench			Avg.	% Gain
Model	Rot.	Ard.	Amg.	All	PR.	Attr.	All	MV	Pers.	D.Rot	D.Tr	All	Avg.	% Gain
Training-Scaled Spatial Model
SenseNova-SI-1.5-InternVL3-8B	92.5	90.8	94.5	93.24	47.1	37.7	45.2	63.9	52.0	43.6	41.7	43.0	59.5	-
Model-Scaled Spatial Agent
GCA (Qwen3-VL-235B-A22B-Thinking) †	82.0	61.8	59.8	64.2	52.8	45.0	51.2	-	58.6	-	-	-	-	-
Vantage on base models

Table 1: Main results on five benchmarks. For each base model, the better result is shown in bold . The final two columns report the average accuracy across all five benchmarks and the relative improvement. † Results are reported as in their original papers.

Model	Question Analysis	View Planning	MindCube-tiny		MMSI-Bench		BLINK		OmniSpatial
Model	Question Analysis	View Planning	Acc.	$\Delta$	Acc.	$\Delta$	Acc.	$\Delta$	Acc.	$\Delta$
Qwen3-VL-4B	–	–	26.0	$-12.9$	27.9	$-5.8$	41.4	$-0.7$	42.1	$-5.1$
	Self	Self	38.9	–	33.7	–	42.1	–	47.2	–
	Cross	Self	51.0	$+12.1$	35.3	$+1.6$	37.6	$-4.5$	50.6	$+3.4$
	Cross	Cross	49.9	$+11.0$	36.2	$+2.5$	36.8	$-5.3$	48.1	$+0.9$
Qwen3-VL-8B	–	–	31.3	$-0.3$	30.1	$0.0$	44.4	$-0.7$	43.9	$-1.2$

Table 2: Cross-source composition results. Self denotes steps executed by the evaluated model, while Cross denotes steps replaced by Qwen3.6-27B. Bold indicates the highest accuracy within each model group.

\Delta

denotes the accuracy change relative to model + Vantage.

Table 5

Model	Augmentation			Benchmark			Avg.	Abs $\Delta$	Relative $\Delta$
Model	View	Analysis	Guidance	MindCube-tiny	MMSI-Bench	OmniSpatial	Avg.	Abs $\Delta$	Relative $\Delta$
Qwen3-VL-4B				26.0	27.9	42.1	32.0	–	–
		✓	✓	34.5	32.1	43.1	36.6	+4.6	+14.4
	✓			29.7	32.7	43.5	35.3	+3.3	+10.3
	✓	✓		38.2	34.0	42.4	38.2	+6.2	+19.4
	✓		✓	36.7	31.4	42.1	36.7	+4.7	+14.7

Table 5: Final VQA context ablation. Checkmarks indicate the component provided to the original VQA. Bold and italic denote the best and second-best average accuracy within each model, respectively. The final two columns report the absolute and relative improvements over its baseline.

Panel of Figure 3 — Figure 3: Major failure types of Vantage. Left : Macro-averaged failure analysis. Right : Representative failure cases by type. More detailed visualizations of these representative failure cases are provided in the Appendix.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Action	Transformation	Small	Medium	Large
Rotation	Pan left	$+\Delta\mathrm{yaw}$	$30^{\circ}$	$60^{\circ}$	$90^{\circ}$
	Pan right	$-\Delta\mathrm{yaw}$	$30^{\circ}$	$60^{\circ}$	$90^{\circ}$
	Tilt up	$+\Delta\mathrm{pitch}$	$30^{\circ}$	$60^{\circ}$	$90^{\circ}$
	Tilt down	$-\Delta\mathrm{pitch}$	$30^{\circ}$	$60^{\circ}$	$90^{\circ}$
	Turn around	$\Delta\mathrm{yaw}$	$180^{\circ}$
Translation	Move right	$+\Delta x$	$0.40s$	$0.80s$	$1.40s$

Appendix

Table 6: Action set for view planning. Each semantic action is deterministically mapped to a geometric transformation with three magnitude levels (small, medium, and large), except for turn around , which uses a fixed

180^{\circ}

rotation. Translation magnitudes are expressed in scene unit

s

Figure 4: Effect of gravity alignment on view synthesis. Given the same input views, reference view, and planned camera action, we compare view synthesis using a gravity-aligned G3T reconstruction and a non-gravity-aligned VGGT- $\Omega$ reconstruction. Gravity alignment provides a ground-consistent coordinate frame for executing relative camera actions, while an unaligned reconstruction can cause the realized viewpoint to deviate from the intended ground-relative motion.

Model	FoV	MindCube-tiny	$\Delta$
Qwen3.6-27B	$120^{\circ}$ (ours)	68.8	–
Qwen3.6-27B	$80^{\circ}$	68.1	$-0.7$
Gemma-4-31B	$120^{\circ}$ (ours)	61.9	–
Gemma-4-31B	$80^{\circ}$	59.4	$-2.5$

Appendix

Table 7: Field-of-view ablation on MindCube.

Model	Backend	MindCube-tiny	MMSI-Bench	Avg.
Qwen3-VL-4B	G3T	38.9	33.7	36.3
Qwen3-VL-4B	VGGT- $\Omega$ + GeoCalib	37.4	30.7	34.1
Qwen3-VL-8B	G3T	31.6	30.1	30.9
Qwen3-VL-8B	VGGT- $\Omega$ + GeoCalib	31.6	29.9	30.8
Qwen3.6-27B	G3T	68.8	44.8	56.8
Qwen3.6-27B	VGGT- $\Omega$ + GeoCalib	67.2	42.2	54.7

Appendix

Table 8: Reconstruction backends ablation.

Figure 5: Failure distribution by model. For each model, we annotate 150 incorrect predictions sampled across the three multi-view benchmarks and categorize them into five failure types. Bars indicate the proportion of annotated failures falling into each category.

Figure 6: Qualitative example of Vantage with GPT-5.4. For clarity, we simplify the original model outputs and present only the key intermediate information relevant to the reasoning process.

Figure 7: Qualitative example of Vantage with GPT-5.4.

Figure 8: Qualitative example of Vantage with GPT-5.4.

Figure 9: Qualitative example of Vantage with GPT-5.4.

Figure 10: Incorrect analysis. Qualitative failure case from Qwen3.6-27B.

Figure 11: Incorrect view plan. Qualitative failure case from Qwen3-VL-4B.

Figure 12: Unreliable synthesis. Qualitative failure case from Gemma-4-31B.

Figure 13: Reasoning failure. Qualitative failure case from Qwen3-VL-4B.

Explore similar work

MV-STRIDE: Enabling MLLMs to Master Multi-View Spatial Reasoning via Hierarchical Capability Modeling
Sep 7, 2026Jin Xu, Xiaojian Huang, Zhuodong Luo +6Spatial Reasoning Benchmarks 3D Spatial Reasoning
MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, China · Huawei Noah’s Ark Lab · Hefei University of Technology, China
CrossView Suite: Harnessing Cross-view Spatial Intelligence of MLLMs with Dataset, Model and Benchmark
May 18, 2026Wei Wang, Yuqian Yuan, Tianwei Lin +4VLM Evaluation Multi-View Consistency
Zhejiang University China
OmniView-Space: Reinforcing Spatial Reasoning via Multi-Perspective Spatial Mapping
Jul 1, 2026Xudong Li, Mengdan Zhang, Peixian Chen +7Visual Spatial Reasoning 3D Spatial Reasoning
Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China · Tencent Youtu Lab · Beijing Institute of Technology

Sep 7, 2026, cs.CVJ/K move · Enter open · S save

MV-STRIDE: Enabling MLLMs to Master Multi-View Spatial Reasoning via Hierarchical Capability Modeling

Jin Xu, Xiaojian Huang, Zhuodong Luo, Zhihong Zhang+5

MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, China · Huawei Noah’s Ark Lab · Hefei University of Technology, China

Spatial Reasoning Benchmarks 3D Spatial Reasoning Spatial Reasoning 3D Scene Understanding

Open paper AskPDF arXiv

Despite the rapid progress of Multimodal Large Language Models (MLLMs) in 2D vision-language tasks, robust multi-view spatial reasoning remains a fundamental bottleneck due to the lack of structured 3D cognitive pathways in existing datasets. To address this, we introduce MV-STRIDE, a Multi-View hierarchical SpaTial Reasoning dataset with Interdependent and DEcomposed capabilitiEs. Moving beyond flat data structures, MV-STRIDE explicitly models the dependency relationships between foundational perception, scene understanding, and complex contextual reasoning, providing a coherent learning pathway aligned with human spatial cognition. We develop a systematic QA generation pipeline leveraging diverse 3D scene sources that enforces cross-view dependency constraints to prevent single-view solvability, generating multi-level spatial reasoning tasks supported by cognitively grounded chain-of-thought supervision for complex inference. Extensive evaluations demonstrate that our multi-stage training framework based on our hierarchical dataset achieves state-of-the-art performance across multiple spatial reasoning benchmarks, notably the multi-view oriented MMSI-Bench. Our approach enables MLLMs to maintain robust, 3D-consistent spatial reasoning across diverse viewpoints. The code and dataset are available at https://co1dspring.github.io/MV-STRIDE/.

Abstract

Explore similar work

MV-STRIDE: Enabling MLLMs to Master Multi-View Spatial Reasoning via Hierarchical Capability Modeling

CrossView Suite: Harnessing Cross-view Spatial Intelligence of MLLMs with Dataset, Model and Benchmark

OmniView-Space: Reinforcing Spatial Reasoning via Multi-Perspective Spatial Mapping