Perspective-aware spatial reasoning involves understanding spatial relationships from specific viewpoints-either egocentric (observer-centered) or allocentric (object-centered). While vision-language models (VLMs) perform well in egocentric settings, their performance deteriorates when reasoning from allocentric viewpoints, where spatial relations must be inferred from the perspective of objects within the scene. In this study, we address this underexplored challenge by introducing Symbolic Projective Layout (SymPL), a framework that reformulates allocentric reasoning into symbolic-layout forms that VLMs inherently handle well. By leveraging four key factors-projection, abstraction, bipartition, and localization-SymPL converts allocentric questions into structured symbolic-layout representations. Extensive experiments demonstrate that this reformulation substantially improves performance in both allocentric and egocentric tasks, enhances robustness under visual illusions and multi-view scenarios, and that each component contributes critically to these gains. These results show that SymPL provides an effective and principled approach for addressing complex perspective-aware spatial reasoning.
Figures & tables
Figure 1: SymPL reformulates allocentric questions into symbolic-layout questions using four factors- projection , abstraction , bipartition , and localization -enabling significantly improved spatial reasoning under allocentric settings.
Figure 2: Overview of SymPL framework. SymPL reformulates an allocentric question into a symbolic-layout question through two stages: 1) Spatial Information Extraction and 2) Question Reformulation using four key factors — projection , abstraction , bipartition , and localization .
Figure 3: Partition rule based on spatial reasoning category. Directional comparisons adopt a linear partition, while distance comparisons employ a circular one.
Group
Method
COMFORT#
3DSRBench
left/right
closer
visibility
facing
left/right
visibility
facing
Random
Random
48.75
48.67
47.27
52.33
50.72
50.00
47.69
General Purpose
LLaVA-NeXT [ 29 ]
47.33
51.58
51.80
52.42
37.39
51.89
58.53
LLaVA-OneVision [ 24 ]
47.92
57.83
51.88
50.42
36.82
47.53
61.42
Molmo [ 12 ]
46.58
46.17
53.36
53.58
39.97
50.29
59.39
Qwen2.5-VL [ 1 ]
48.17
72.33
51.17
51.33
36.25
48.40
65.03
Table 1: Quantitative results on allocentric questions. Bold indicates the best, while underline represents the second best results.
Figure 4: Allocentric spatial reasoning examples. Qwen2.5-VL + SoM and APC-Vis exhibited limited allocentric spatial reasoning performance across various categories. In contrast, our SymPL effectively handled allocentric questions by reformulating them into symbolic-layout questions.
Method
COCOSPATIAL
left/right
above/below
Random
50.25
50.00
LLaVa-NeXT
88.67
76.00
LLaVa-OneVision
87.67
82.67
Molmo
81.17
72.17
Qwen2.5-VL
86.33
78.75
Table 2: Quantitative results on egocentric questions. Bold indicates the best, while underline represents the second best results.
Method
COMFORT VI
Allocentric
Egocentric
left/right
front/behind
closer
Random
49.69
51.75
49.13
LLaVa-NeXT
58.19
58.75
89.13
LLaVa-OneVision
72.00
95.88
98.75
Molmo
44.56
93.88
74.50
Table 3: Quantitative results on perspective-aware reasoning under visual illusions. Bold indicates the best, while underline represents the second best results.
Method
left/right
closer
visibility
facing
Qwen2.5-VL
67.50
70.50
58.50
57.50
Qwen2.5-VL + CoT
57.50
66.00
60.50
50.50
Qwen2.5-VL + SoM
72.50
51.50
61.50
54.50
Qwen2.5-VL + SCAFFOLD
66.50
53.50
54.00
55.00
APC-Num
44.50
28.50
55.00
23.50
APC-Vis
53.00
31.50
61.50
16.50
Table 4: Quantitative results on viewpoint-aware consistency across multiple views. Bold indicates the best, while underline represents the second best results.
Figure 5: Ablation results of each key factor. (a) projection , (b) abstraction , (c) bipartition , (d) localization . The darker bar indicates the configuration used in SymPL.
Setting
P
A
B
L
left/right
closer
visibility
facing
1
46.60
63.80
52.00
52.80
2
✓
89.20
64.80
51.20
52.00
3
✓
✓
96.40
81.00
90.80
100.00
4
✓
✓
✓
97.00
91.00
84.60
100.00
5 (Ours)
✓
✓
✓
✓
100.00
100.00
100.00
100.00
Table 5: Ablation results on the effectiveness of four key factors: P rojection, A bstraction, B ipartition, and L ocalization. Results show the average success rate of five general-purpose VLMs for each category: left/right , closer , visibility , and facing .
Figure 6: Error breakdown in SymPL’s reasoning pipeline.
Spatial reasoning studies how systems represent geometry, relations, viewpoints, and motion to understand scenes, plan actions, and solve spatial tasks.
Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Pengcheng Laboratory, Shenzhen, China · School of Software Engineering, Chongqing University, Chongqing, China