ChainSpace: A Chained-Reasoning Paradigm for Spatial Intelligence

Xiaohan Zhang; Feng Gu; Xudong Rao; Xuhao Pan; Tao Wei; Zhou Pan; Kun Zhan

cs.CVAug 16, 2026

ChainSpace: A Chained-Reasoning Paradigm for Spatial Intelligence

Authors: Xiaohan Zhang, Feng Gu, Xudong Rao, Xuhao Pan, Tao Wei, Zhou Pan, Kun Zhan

Organizations: College of Information Science and Electronic Engineering, Zhejiang University · Li Auto Inc.

Abstract

Spatial intelligence requires foundation models to maintain coherent spatial state across interactions with the physical world. However, existing data-centric approaches typically treat spatial reasoning as independent question-answer instances, enabling shortcut-based answering and providing limited supervision for persistent spatial understanding. To address this, we introduce ChainSpace, a chained-reasoning paradigm that structures spatial reasoning as a state-preserving multi-round process. In this paradigm, spatial questions are organized into logically constrained and jointly consistent chains, where later questions depend on spatial constraints established in earlier rounds. Following this principle, we instantiate ChainSpace-Bench, a manually annotated real-world multi-round benchmark with a Chain-Aware Metric, and ChainSpace-Pipeline, a simulator-based chain-structured supervision generation framework for spatial intelligence training. Experiments show that ChainSpace-Bench exposes chain-level failures that are not captured by isolated question accuracy. Additionally, with a relatively small amount of simulator-generated chained data, models trained by ChainSpace-Pipeline achieve the best performance among open-source models on ChainSpace-Bench and transfer competitively to multiple external spatial intelligence benchmarks. These results establish ChainSpace as an effective paradigm for more faithful evaluation and more data-efficient learning of spatial intelligence.

Figures & tables

Figure 1: Illustration of the ChainSpace paradigm. Given a video trajectory, spatial questions are organized into a logically constrained chain, where later questions depend on spatial constraints established in earlier rounds. (a) A model following an independent-QA behavior may answer an early question correctly using local cues from a single frame, but fail on later questions that require preserving and updating spatial state. This illustrates why isolated question accuracy can overestimate spatial intelligence. (b) By reasoning with the preserved spatial state across frames and question turns, the model can produce answers that remain jointly consistent throughout the chain.

Figure 2: Illustration of our task templates. Each trajectory is instantiated into four 3-round question templates, i.e. , object count , appearance order , relative direction , and route plan . All templates follow the same chained design principle, where later questions depend on the spatial constraints established by earlier ones. For object count , the chain goes from global counting to room-wise decomposition and then to finer-grained counting within a specific room. For appearance order , it goes from earliest object identification to first-appearance room localization and then to object-location reasoning within that room. For relative direction , it goes from direction perception to hypothetical viewpoint change and then to action prediction. For route plan , it goes from room visitation sequence to room adjacency and then to shortest-path reasoning.

Figure 3: Frame count distribution of ChainSpace-Bench . The benchmark covers trajectories with diverse lengths, with most samples concentrated in the moderate range and a long tail of substantially longer videos.

Figure 4: The A/B/C/D answer distribution in ChainSpace-Bench by turn. The option labels are reasonably balanced across rounds, reducing the possibility that models can perform well by exploiting simple answer-frequency bias.

Model	Avg. Question Accuracy	Avg. Chain-Aware Score
Proprietary Models
Seed-2.0-Pro [ 6 ]	0.4289	0.2722
GPT-5 [ 25 ]	0.4825	0.3262
Gemini-3.1-Pro [ 16 ]	0.3453	0.1882
Open-source General Models
Qwen2.5-VL-7B-Instruct [ 3 ]	0.2344	0.0819

Table 1: Results on ChainSpace-Bench . Open-source models and our method are all built on Qwen-family (Qwen2.5-VL-7B-Instruct [ 3 ] and Qwen3-VL-8B-Instruct [ 40 ] ) for fair comparison. Existing baselines are trained with different data scales, including CamBrain-S (

\sim

590K), RoboBrain2.0 (

\sim

5.3M), RoboBrain2.5 (

\sim

12.4M), and SenseNova-SI (

\sim

8M), while our ChainSpace-Pipeline uses only

\sim

50K simulator-generated samples. The text in red and blue denotes the best and second-best results among open-source models.

Metric	( Ours ) ChainSpace-Pipeline	SenseNova-SI
( Chain-level metrics ) Full-chain Exact Match	+0.0153 [0.0054, 0.0261] (0.0044)	+0.0038 [-0.0100, 0.0176] (0.6380)
( Chain-level metrics ) Avg. Chain-Aware Score	+0.0165 [0.0079, 0.0256] (<0.0001)	+0.0033 [-0.0086, 0.0154] (0.5980)
( Turn-level metrics ) Turn-1 Accuracy	+0.0261 [0.0146, 0.0383] (<0.0001)	+0.0092 [-0.0146, 0.0330] (0.4720)
( Turn-level metrics ) Turn-2 Accuracy	+0.0107 [-0.0038, 0.0253] (0.1692)	-0.0215 [-0.0376, -0.0054] (0.0116)
( Turn-level metrics ) Turn-3 Accuracy	+0.0130 [-0.0077, 0.0345] (0.2416)	+0.0345 [0.0084, 0.0606] (0.0108)

Table 2: Paired comparison against the same baseline (Qwen3-VL-8B-Instruct [ 40 ] ) on ChainSpace-Bench . Each cell reports the mean paired difference

\Delta

, the 95% paired-bootstrap confidence interval, and the paired-bootstrap

p

-value in the format

\Delta

[CI] (

p

). Bold indicates statistically significant change (

p<0.05

Model	Setting	Avg. Acc.	Q2 Acc.	Q3 Acc.	Chain-Aware
Qwen3-VL-8B [ 40 ]	Pred	0.3344	0.3574	0.3497	0.1522
	Gold	0.3563	0.3635	0.4095	0.1589
	Gain	+6.55%	+1.71%	+17.10%	+4.40%
SenseNova-SI-8B [ 7 ]	Pred	0.3418	0.3359	0.3842	0.1555
	Gold	0.3635	0.3528	0.4325	0.1628
	Gain	+6.35%	+5.03%	+12.57%	+4.69%

Table 3: Oracle-history intervention on ChainSpace-Bench. “Pred” denotes normal inference using model-generated previous answers, while “Gold” replaces previous answers with ground-truth answers during inference. Relative gains are computed as

(\mathrm{Gold}-\mathrm{Pred})/\mathrm{Pred}

. The text in red and blue denotes the best and second-best results of each setting.

Model	Avg. Question Accuracy	Chain-Aware Score
Qwen3-VL-8B-Instruct [ 40 ]	0.3348	0.1531
SenseNova-SI-8B [ 7 ]	0.3427	0.1562
ChainSpace-Pipeline-8B (Ours)	0.3546	0.1690

Table 4: Performance on the enlarged ChainSpace-Bench containing 800 videos and 3,200 chains. The red and blue denote the best and second-best results, respectively.

Models	ChainSpace-Bench		VSI-Bench [ 41 ]	MMSI-Bench [ 43 ]	MindCube [ 34 ]
Metric	Avg. Q. Acc	Avg. Chain-Aware Score	MRA, Acc	Acc	Acc
Qwen3-VL-8B-Instruct [ 40 ]	0.3344	0.1522	57.9	31.1	29.4
Independent-QA Control (8B)	0.3236	0.1482	58.1	30.4	30.2
ChainSpace-Pipeline-8B (Ours)	0.3510	0.1687	58.8	31.8	31.1

Table 5: Controlled comparison of different training paradigms on ChainSpace-Bench and external spatial reasoning benchmarks. All settings use the same backbone. Independent-QA uses the same videos, questions, answers, supervision content, and training configuration as ChainSpace-Pipeline, while removing cross-round dependencies. The red and blue denote the best and second-best results, respectively.

Models	VSI-Bench [ 41 ]	MMSI-Bench [ 43 ]	MindCube* [ 34 ]	ViewSpatial [ 19 ]	EmbSpatial [ 15 ]
Metric	MRA, Acc	Acc	Acc	Acc	Acc
Random Choice	34.0	25.0	33.0	26.3	-
Proprietary Models
Seed-1.6-2025-06-15 [ 5 ]	49.9	38.3	48.7	43.8	-
Gemini-2.5-Pro-2025-06 [ 12 ]	53.5	38.0	57.6	46.0	78.9
Grok-4-2025-07-09 [ 39 ]	47.9	37.8	63.5	43.2	75.7

Table 6: Comparison between proprietary models, open-source general models, and open-source spatial intelligence models across five spatial intelligence benchmarks. We report our 50K and 150K settings to highlight data efficiency. MindCube* denotes MindCube-Tiny.

Model	Overall	Cnt./Exist.	Obj. Loc.	3D Info.	MV/XImg.	Spatial Rel.	Move/Nav.
Baseline	43.54	55.99	49.11	56.00	23.05	66.13	21.32
SenseNova-SI-150K [ 7 ]	45.35	57.36	51.91	58.73	24.75	68.45	19.51
Ours-50K	43.94	56.97	47.07	56.78	23.52	66.79	23.74
Ours-50K+SenseNova-SI-100K [ 7 ]	45.51	58.74	58.85	58.34	25.78	68.61	18.90

Table 7: Comparison of Qwen3-VL-8B-Instruct [ 40 ] trained with different data source under Chance-Adjusted Accuracy (CAA, %) on the SITE-Bench [ 35 ] . The text in red and blue denotes the best and second-best results.

Model	Method	Vanilla	Flip	Circ	Flip+Circ
Qwen3-VL-8B-Instruct [ 40 ]	Baseline	0.566	0.498	0.533	0.475
Qwen3-VL-8B-Instruct [ 40 ]	Ours-50K	0.564	0.499	0.541	0.483
InternVL3-8B [ 49 ]	Baseline	0.524	0.444	0.396	0.324
InternVL3-8B [ 49 ]	Ours-150K	0.541	0.466	0.447	0.384

Table 8: Performance comparison on 3DSRBench [ 22 ] under different evaluation settings. Bold indicates the better result within each model group.

Benchmark	Qwen3-VL-8B-Instruct [ 40 ]	ChainSpace-Pipeline-8B (Ours)	Representative Task-Level Gains
VSI-Bench [ 41 ]	57.90	58.82	Route Planning $+3.56$ ; Absolute Distance $+1.54$ ; Relative Distance $+1.30$ ; Appearance Order $+1.18$
MMSI-Bench [ 43 ]	31.10	31.80	Camera–Camera Relation $+6.41$ ; Attribute Measurement $+6.25$ ; Appearance $+2.98$ ; Multi-step Spatial Reasoning $+2.02$
ViewSpatial-Bench [ 19 ]	42.20	42.96	Person-perspective Object Orientation $+3.30$ ; Scene Simulation $+0.85$
3DSRBench (Circ.) [ 22 ]	53.30	54.10	Viewpoint-to-Object Relation $+2.64$ ; Next-to Relation $+1.79$ ; In-front-of Relation $+1.15$

Table 9: Task-level analysis on external spatial reasoning benchmarks. ChainSpace-Pipeline-8B yields larger gains on compositional, viewpoint-conditioned, and planning-related spatial tasks.

Model	Easy		Medium		Hard		Exec. Avg.
	SR	SPL	SR	SPL	SR	SPL
GPT-4o [ 24 ]	67.36	54.31	41.67	35.71	27.78	21.15	41.33
Qwen2.5-VL-7B-Instruct [ 3 ]	41.67	32.55	22.92	17.43	10.42	5.67	21.78
Qwen3-VL-8B-Instruct [ 40 ]	50.69	38.04	42.36	34.05	26.39	20.87	35.40
SenseNova-SI-8B [ 7 ]	46.53	33.48	25.00	17.64	22.22	15.32	26.70
ChainSpace-Pipeline-8B (Ours)	53.47	39.65	42.36	32.66	29.86	22.82	36.80

Table 10: End-to-end embodied navigation performance on NavBench [ 28 ] . All models are evaluated without NavBench-specific training. Red and blue denote the best and second-best results among open-source models, respectively. Exec. Avg. denotes the average of SR and SPL across the three difficulty levels.

Setting	$\boldsymbol{\alpha_{1},\alpha_{2},\alpha_{3}}$	$\boldsymbol{\beta_{12},\beta_{23},\beta_{13}}$	$\boldsymbol{\gamma}$	Qwen3-VL-8B-Instruct [ 40 ]	RoboBrain2.5 [ 30 ]	SenseNova-SI [ 7 ]	Ours
$W_{1}$	(0.5,0.5,0.5)	(2.0,1.5,1.25)	3.0	0.1522	0.1339	0.1555	0.1687
$W_{2}$	(1.0,1.0,1.0)	(0.0,0.0,0.0)	0.0	0.3343	0.3384	0.3418	0.3510
$W_{3}$	(0.0,0.0,0.0)	(1.0,1.0,1.0)	2.0	0.1161	0.0925	0.1189	0.1327
$W_{4}$	(1.0,1.0,1.0)	(0.0,0.0,0.0)	2.0	0.2298	0.2221	0.2357	0.2459
$W_{5}$	(1.0,1.0,1.0)	(1.0,1.0,1.0)	1.0	0.1979	0.1847	0.2025	0.2145

Table 11: Ranking robustness under different weight settings of our Chain-Aware Metric (Eq. ( 4 )) on ChainSpace-Bench . We vary the weights of the first-order terms

(\alpha_{t})

, second-order terms

(\beta_{ij})

, and the third-order term

(\gamma)

, and report the resulting normalized Avg. Chain-Aware scores for representative methods in Table 1 . The text in red and blue denotes the best and second-best results.

Metric	25K	50K	75K	125K
Full-chain Exact Match	+0.0038 (0.3900)	+0.0153 ∗ (0.0044)	+0.0130 ∗ (0.0220)	+0.0153 ∗ (0.0153)
Avg. Chain-Aware Score	+0.0038 (0.2740)	+0.0165 ∗ (<0.0001)	+0.0141 ∗ (0.0032)	+0.0141 ∗ (0.0048)

Table 12: Scaling behavior of chained supervision on ChainSpace-Bench . We report the paired improvement

\Delta

over the same untuned baseline under different training data scales. Each entry is shown as

\Delta

followed by the paired-bootstrap

p

-value in parentheses. ∗ indicates statistical significance (

p<0.05

and 95% CI excluding 0). The text in red and blue denotes the best and second-best results.

Figure 5: Visualization of qualitative examples on ChainSpace-Bench . SenseNova-SI [ 7 ] answers some individual turns correctly, whereas our model maintains consistent answers across all rounds.

Figure 6: Visualization of counterfactual query examples. With the same video context but a changed room-pair question, our model remains correct while SenseNova-SI [ 7 ] fails, suggesting a more complete maintained spatial state.

Figure 7: Visualization of a representative failure case. The model confuses the complete room order in a many-room trajectory, while preserving correct local spatial relations for later reasoning.

Spatial Reasoning

Spatial reasoning studies how systems represent geometry, relations, viewpoints, and motion to understand scenes, plan actions, and solve spatial tasks.

See all 154 papers

Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests
Jul 24, 2026Patrick Rim, Tom Long, Ekta Prashnani +63D Spatial Reasoning Visual Spatial Reasoning
1NVIDIA Research · 2Yale University
SpatialChain: A Benchmark for Auditing Spatial Reasoning Faithfulness in VLMs
Oct 5, 2026Rafael Teixeira Sousa, Vinícius Paulo Lopes de Oliveira, Elisa Ayumi Masasi de Oliveira +5VLM Evaluation Spatial Reasoning Benchmarks
Universidade Federal de Mato Grosso (UFMT) · Universidade Federal de Goiás (UFG) · Advanced Knowledge Center for Immersive Technologies (AKCIT)
Ouroboros-Spatial: Closing the Data-Model Loop for Spatial Reasoning
Jun 10, 2026Enhan Zhao, Wei Wu, Yuanrui Zhang +23D Spatial Reasoning Synthetic Data Generation
1Peking University · 2Ant International · 3The University of Hong Kong

Jul 24, 2026J/K move · Enter open · S save

Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests

Patrick Rim, Tom Long, Ekta Prashnani, Ruth Rosenholtz+5

1NVIDIA Research · 2Yale University

3D Spatial Reasoning Visual Spatial Reasoning Object Counting Multimodal Model Evaluation Spatial Reasoning Benchmarks Spatial Reasoning in Language Models+1

Open paper AskPDF arXiv

Multimodal large language models (MLLMs) excel at visual interpretation but fail on spatial reasoning tasks that humans solve reliably. Existing benchmarks evaluate these models as black boxes, limiting their ability to identify the underlying causes of lower performance: when a model fails a spatial reasoning task, it remains difficult to ascertain whether the hurdle is perceptual, such as recognizing object boundaries, or cognitive, such as reasoning about occlusion to infer hidden geometry. We introduce Spatial-IQ, a hierarchical diagnostic framework that decomposes object counting in stacked 3D structures into 9 perceptual and cognitive sub-tasks organized by the developmental stages of human spatial cognition, with mental rotation as an additional target probe. Using NVIDIA Isaac Sim, we procedurally generated a diverse dataset of roughly 80,000 stacked 3D structures with per-task ground truth. We evaluate models across three output formats (free-response text, multiple-choice images, and image editing) alongside a human baseline. The Spatial-IQ framework shows that top-performing models often succeed at the target task (object counting) without succeeding on the lower-level sub-tasks intended to support it, and that models differ in how much of these hierarchical chains they preserve, often revealing shortcut behavior that raw target-task accuracy alone would obscure. Finally, we demonstrate that training models with chain-of-thought (CoT) supervision over our hierarchical sub-tasks, combined with reinforcement learning with verifiable rewards, significantly improves both spatial consistency across sub-tasks and target-task accuracy, supporting the value of the proposed decomposition as both a diagnostic tool and a training signal.

Abstract

Spatial Reasoning

Explore similar work

Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests

SpatialChain: A Benchmark for Auditing Spatial Reasoning Faithfulness in VLMs

Ouroboros-Spatial: Closing the Data-Model Loop for Spatial Reasoning