Spatial reasoning over three-dimensional scenes is a core capability for embodied intelligence, yet continuous model improvement remains bottlenecked by the cost of geometric annotation. The self-evolving paradigm offers a promising path, but its reliance on model consensus to construct pseudo-labels causes training to reinforce rather than correct the model's own geometric errors. We identify a property unique to 3D spatial reasoning that circumvents this limitation: ground truth is a deterministic consequence of the underlying geometry, computable exactly from point clouds and camera poses without any model involvement. Building on this insight, we present SpatialEvo, a self-evolving framework for 3D spatial reasoning, centered on the Deterministic Geometric Environment (DGE). The DGE formalizes 16 spatial reasoning task categories under explicit geometric validation rules and converts unannotated 3D scenes into zero-noise interactive oracles, replacing model consensus with objective physical feedback. A single shared-parameter policy co-evolves across questioner and solver roles under DGE constraints: the questioner generates physically valid spatial questions grounded in scene observations, while the solver derives precise answers against DGE-verified ground truth. A task-adaptive scheduler endogenously concentrates training on the model's weakest categories, producing a dynamic curriculum without manual design. Experiments across nine benchmarks demonstrate that SpatialEvo achieves the highest average score at both 3B and 7B scales, with consistent gains on spatial reasoning benchmarks and no degradation on general visual understanding.
Figures & tables
Figure 1: Comparison of three training paradigms. Static Data Tuning relies on manually annotated fixed datasets. Consensus Self-Evolve eliminates manual annotation but introduces biased ground truth via model voting. SpatialEvo (Ours) leverages the DGE to compute exact ground truth from 3D scene assets, enabling zero-noise dynamic curriculum learning.
Figure 2: Overview of SpatialEvo . The DGE computes zero-noise ground truth directly from 3D point clouds and camera pose sequences, enabling the Spatial Self-Evolution Loop to jointly optimize a single VLM as both Questioner and Solver via Spatial-Grounded GRPO.
Benchmark
Qwen2.5-VL-3B
Qwen2.5-VL-7B
Baseline
SpatialLadder
SpaceR
SpatialSSRL
SpatialEvo
Baseline
VILASR
SpaceR
SpatialSSRL
SpatialEvo
VSI-Bench ( Yang et al., 2025a )
28.1
45.7
36.0
28.0
39.2
31.1
45.4
36.8
33.7
46.1
RealWorldQA ( Zhang et al., 2024 )
63.4
57.1
61.4
65.4
66.5
69.5
57.9
64.7
69.9
66.7
EmbSpatial ( Du et al., 2024 )
55.9
57.6
55.6
59.8
61.2
63.6
47.8
60.3
69.3
66.0
SpatialViz ( Wang et al., )
24.2
28.6
31.9
25.9
25.4
27.0
29.8
30.9
28.4
28.6
STARE ( Li et al., 2025c )
33.1
26.4
36.8
36.8
36.9
41.8
21.4
36.2
43.3
41.3
Table 1: Main Results. Nine benchmarks spanning spatial reasoning and general visual understanding. Shaded columns indicate SpatialEvo (ours). Bold denotes the best result per benchmark under each backbone.
Variant
VSI-Bench
RealWorldQA
EmbSpatial
SpatialViz
STARE
CoreCognition
ViewSpatial
V-STAR
MMStar
Avg
Δ Avg
SpatialEvo (Ours)
46.1
66.7
66.0
28.6
41.3
60.2
43.2
78.0
62.5
54.7
–
Architecture Design
w/o Questioner
40.2
67.1
65.7
27.7
39.0
60.5
40.4
77.0
59.9
53.1
↓ 1.6
w/o Solver
36.6
70.2
61.8
26.1
39.8
58.5
34.5
75.4
60.5
51.5
↓ 3.2
w/o Physical Grounding
18.8
68.5
63.5
23.4
39.7
59.7
35.4
77.0
60.6
49.6
↓ 5.1
w/o Adaptive Scheduler
43.4
68.5
68.0
27.5
39.5
60.3
43.2
77.0
62.4
54.4
↓ 0.3
Table 2: Ablation study on SpatialEvo (Qwen2.5-VL-7B). Δ Avg denotes performance drop relative to the full model. Bold denotes the best Avg.
Paradigm
Method
Numerical Question
Multiple-Choice Question
Avg.
Obj. Count
Abs. Dist.
Obj. Size
Room Size
Rel. Dist.
Rel. Dir.
Route Plan
Appr. Order
Qwen2.5-VL-3B (Base)
33.5
21.1
17.9
22.6
32.8
42.7
29.9
21.0
28.0
RL
w/ SpatialLadder RL
62.8
29.8
59.2
27.8
38.9
44.2
33.5
24.9
40.1
SpatialEvo (Online RL)
65.2
35.1
61.3
51.4
46.2
44.3
29.9
26.5
46.3
SFT
w/ SpatialLadder Data
63.0
31.9
61.2
43.0
43.0
43.6
32.5
31.7
43.7
w/ SpaceR Data
28.3
25.9
36.3
34.9
35.1
46.8
35.1
43.4
36.3
Table 3: Comparison of online self-evolution and static learning paradigms on VSIBench. SpatialEvo outperforms static RL and SFT baselines across both training paradigms.
Figure 3: Training dynamics of SpatialEvo. Questioner rewards stabilize as valid question generation is learned (Left); Solver accuracy improves while the invalid ratio declines (Middle); the Adaptive Scheduler up-weights harder categories (Rel. Dir., Rel. Dist.) and down-weights easier ones (Right).
Method
Numerical Question
Multiple-Choice Question
Avg.
Obj. Count
Abs. Dist.
Obj. Size
Room Size
Rel. Dist.
Rel. Dir.
Route Plan
Appr. Order
Qwen2.5-VL-7B (Base)
30.7
10.2
37.14
43.40
37.2
37.1
32.0
27.8
31.1
SpatialEvo
Iter 1
60.6
28.9
60.3
42.5
42.7
44.4
32.0
30.7
44.2
Iter 2
63.3
28.9
62.5
33.3
43.4
43.8
30.9
36.7
45.0
Iter 3
65.0
26.9
63.4
47.3
43.0
43.4
28.9
32.9
45.1
Table 4: Effect of the Adaptive Scheduler across iterative self-evolution on VSI-Bench subtasks. Both groups are initialized from the same base model and trained for four iterations; the only difference is whether the Adaptive Scheduler is enabled.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Table 8
Input Modality
ScanNet
ScanNet++
ARKitScenes
Total
Scene-level
846
171
620
1,637
Image-pair
884
298
728
1,910
Single-image
281
97
440
818
Total
2,011
566
1,788
4,365
Appendix
Table 10: Composition of the online RL context pool by modality and data source.
Figure 4: Evolution of the Questioner across training steps on the single-view relative direction task. Observations become increasingly detailed and spatially grounded, with generated questions more tightly coupled to visual evidence.
Figure 5: Evolution of the Questioner across training steps on the single-view relative direction task. Observations become increasingly detailed and spatially grounded, with generated questions more tightly coupled to visual evidence.
3D spatial reasoning studies how AI systems infer geometry, object relations, viewpoints, and motion to understand and act in three-dimensional scenes.