When solving complex spatial reasoning tasks, ordinary humans naturally revisit the scenes from task-relevant viewpoints and mentally reconstruct the spatial layouts from those viewpoints. But can Large Vision-Language Models (LVLMs) proactively think in the same way? We term this ability Active Spatial Mentalization (ASM), an intrinsic component of visual intelligence that remains substantially underdeveloped in current LVLMs. To bridge this gap, we introduce SpatialASM, a scalable framework to internalize ASM into LVLMs through training-time spatial simulation. SpatialASM employs an allocentric novel-view generator to ground hypothetical viewpoints in visual evidence and produces two complementary forms of supervision: i) explicit ASM trajectories that teach LVLMs when and how to adaptively invoke view planning and imagination during spatial reasoning, and ii) automatically-constructed perspective-taking tasks that further strengthens viewpoint-conditioned scene imagination. On seven benchmarks spanning relatively perception-oriented spatial understanding to more challenging tasks requiring multi-image, cross-view and viewpoint-transformation reasoning, SpatialASM yields performance gain on LVLMs of four LVLM backbones in spatial reasoning while retaining their original general vision capabilities. The experimental results demonstrate that our approach improves average accuracy across the seven spatial benchmarks. And the synthesis of ASM trajectories provides a scalable pathway to alleviate the scarcity of spatial Chain-of-Thought (CoT) supervision for eliciting active viewpoint exploration and mental scene simulation.
Figures & tables
Figure 1: SpatialASM improves LVLM spatial reasoning by eliciting active exploration and imagination. (a) The qualitative example shows that the baseline LVLM ( red ) fails to associate objects across views, whereas our model ( green ) uses the internal viewpoint-planning skill and reasons over the imagined observation to reach the correct answer. (b) Quantitative experiments demonstrate that SpatialASM delivers consistent performance gains on seven benchmarks.
Figure 2: (a) Prior data-centric approaches mostly provide answer-only supervision. (b) Reasoning with external calls of world models adds unnecessary intermediate observations and extra inference time. (c) SpatialASM enables LVLMs to intrinsically produce plan-imagine-reason CoT.
Figure 3: Overview of SpatialASM. (1) An allocentric novel-view generator (NVG, G ) renders the scene from a queried viewpoint, and GeoP measures how much a rendered view perturbs the 3D reconstruction of the inputs. (2) An LVLM plans viewpoints, G renders them, and filtered trajectories are converted to captions for SFT. (3) GeoP-filtered candidate views form verifiable perspective-taking questions for RL.
Figure 4: Evaluation of external ASM . We evaluate vanilla VLM reasoning, and two external ASM strategies with different viewpoint-planning methods. Both planning-based settings invoke SpatialASM-NVG and condition the final reasoning on the original and generated observations.
Models
Spatial Reasoning
Gnr.-VQA
Mindcube
MMSI
Spa4576d
Viewspa
3DSR
EmbSpa
ESI
Avg.
Avg.
Representative Spatial Models
SpaceR-7B
36.6
28.8
45.1
38.4
54.5
37.3
40.2
40.1
59.4
VST-RL-7B
39.3
35.1
43.4
47.0
60.1
72.0
40.1
48.1
62.1
HoliSpatial-8B
52.0
30.6
30.5
50.1
49.7
73.8
41.8
46.9
63.1
SenseNova-8B
50.4
29.6
41.3
46.4
55.2
70.8
36.5
47.2
63.7
Table 1: Performance comparison on 7 spatial reasoning and 3 general-VQA benchmarks. We report the average accuracy over all benchmarks within each category. The performance of several spatial models are also provided (7B variants on Qwen2.5-VL-7B; 8B variants on Qwen3-VL-8B).
Subtask
Qwen3-VL-8B
SpatialASM-8B
Vanilla
+ External ASM
Vanilla
+ External ASM
Accuracy
T(s)
Accuracy
TNVG (s)
Accuracy
T(s)
Accuracy
TNVG (s)
MindcubeAmong
33.0
12.6
42.0
57.6
39.5
18.6
45.0
56.8
MindcubeAround
52.0
7.5
45.5
38.1
58.0
7.2
61.5
46.8
Table 2: Accuracy and average inference time of different methods on benchmark subtasks. (i) Vanilla: the LVLM reasons directly from the input observations alone, and the reasoning duration is reported; (ii) + External ASM: Similar to Sec. 4.1 , the LVLM externally invokes SpatialASM-NVG to help reasoning. The time to synthesize novel-view observations is reported.
Table 7
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Input Length
∣I∣=1
∣I∣=2
∣I∣=3
∣I∣=4
Total
Size
35,026
34,835
34,846
35,023
139,730
Appendix
Table 5: The samples numbers of each input-image length.
Figure 5: Examples in SpatialASM-edit140k. (Case 1, 2: ∣I∣=1 , Case 3: ∣I∣=2 )
Data Source
Holi-Spatial
VST-500k
Mindcube-train
Total
Sample No.
91,648
34,741
1,965
128,354
Appendix
Table 6: The number of samples from each data source in SpatialASM-128k
Figure 6: Example in SpatialASM-128k. Note that Iedit, 1 does not exist in the ASM trajectory.
Figure 7: Examples of view-imagination RL tasks.
Models
MMMUVAL
MMStar
RealWorldQA
Average
Representative Spatial Models
SpaceR-7B
51.8
61.4
65.0
59.4
VST-RL-7B
51.2
63.7
71.4
62.1
HoliSpatial-8B
56.0
65.4
68.0
63.1
SenseNova-8B
55.1
64.5
71.6
63.7
Baselines & Our Model (Qwen3-VL-4B)
Appendix
Table 7: Performance comparison on three general-VQA benchmarks.
Figure 8: Examples of novel-view observations rendered by SpatialASM-NVG.
Figure 9: Case 1: Observations generated by different novel-view generators.
Figure 10: Case 2: Observations generated by different novel-view generators.
Figure 11: Qualitative comparison of Qwen3-VL-8B and SpatialASM-8B.
Figure 12: Qualitative comparison of Qwen3-VL-8B and SpatialASM-8B.
Figure 13: Qualitative comparison of Qwen3-VL-8B and SpatialASM-8B.
Figure 14: Qualitative comparison of Qwen3-VL-8B and SpatialASM-8B.