Paper deep dive
SpatialBlock: Enhancing Spatial Intelligence in LVLMs via Synthetic Block-Stacking Problem
Soohyun Ryu, Sohee Kim, Eunho Yang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/13/2026, 6:15:31 AM
Summary
The paper introduces SpatialBlock-15k, a synthetic dataset of 15,000 block-stacking problems designed to enhance the spatial intelligence of Large Vision-Language Models (LVLMs). The dataset covers 3D-to-2D projection, viewpoint transformation, and structural combination, utilizing controlled color modulation as visual cues. Experiments show that LVLMs trained on this dataset significantly outperform baselines in spatial reasoning tasks.
Entities (12)
Relation Signals (9)
SpatialBlock-15k ā containstask ā 3D-to-2D projection
confidence 95% Ā· The dataset comprises three categories of spatial reasoning tasks: (1) 3D-to-2D projection...
SpatialBlock-15k ā containstask ā viewpoint transformation
confidence 95% Ā· The dataset comprises three categories of spatial reasoning tasks: ... (2) viewpoint transformation...
SpatialBlock-15k ā containstask ā structural combination
confidence 95% Ā· The dataset comprises three categories of spatial reasoning tasks: ... and (3) structural combination.
SpatialBlock-3B-direct ā trainedon ā SpatialBlock-15k
confidence 95% Ā· SpatialBlock-3B-direct (Ours) 15K
SpatialBlock-7B-reason ā trainedon ā SpatialBlock-15k
confidence 95% Ā· SpatialBlock-7B-reason (Ours) 15K
SpatialBlock-3B-direct ā evaluatedon ā SB-Bench
confidence 90% Ā· Table 1: Main Results... SpatialBlock-3B-direct (Ours) ... SB-Bench 94.8
SpatialBlock-7B-reason ā evaluatedon ā MindCube
confidence 90% Ā· Table 1: Main Results... SpatialBlock-7B-reason (Ours) ... MindCube 37.9
SpatialBlock-7B-reason ā usestrainingmethod ā GRPO
confidence 90% Ā· SpatialBlock-reason refers to the model that performs reasoning before producing the final answer... optimize the model using reinforcement learning via Group Relative Policy Optimization (GRPO).
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large Vision-Language Models (LVLMs) have achieved strong performance on diverse visual tasks, yet their ability to reconstruct and reason about the 3D structure of the scene depicted in 2D images -- referred to as spatial intelligence -- remains limited. Existing approaches attempt to address this gap by using real-scene spatial question answering datasets that require dense geometric annotations. However, constructing such labels is costly, time-consuming, and often noisy due to reliance on external perception modules. In this work, we propose a novel paradigm inspired by human cognitive development: learning foundational spatial skills through structured block-manipulation tasks. We introduce SpatialBlock-15k, a synthetic dataset of 15,000 block-stacking problems covering 3D-to-2D projection, viewpoint transformation, and structural combination. The dataset further incorporates controlled color modulation as visual cues to encourage anchor-based reasoning in visually complex conditions. Experiments demonstrate that LVLMs trained on our dataset through either direct answering or reasoning-based prediction significantly outperform baselines and generalize to real-world spatial tasks, despite the dataset's synthetic and compact nature. Code and data are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2609.07064v1
- Canonical: https://arxiv.org/abs/2609.07064v1
Trouble viewing inline? Open PDF directly ā
Full Text
71,044 characters extracted from source content.
Expand or collapse full text
SpatialBlock: Enhancing Spatial Intel- ligence in LVLMs via Synthetic Block- Stacking Problem Soohyun Ryu ā & Sohee Kim ā KAIST South Korea rsoohyun, joyhee@kaist.ac.kr Eunho Yang ā KAIST & AITRICS South Korea yangeh@kaist.ac.kr Abstract Large Vision-Language Models (LVLMs) have achieved strong performance on diverse visual tasks, yet their ability to reconstruct and reason about the 3D structure of the scene depicted in 2D images ā referred to as spa- tial intelligence ā remains limited. Existing approaches attempt to address this gap by using real-scene spatial question answering datasets that re- quire dense geometric annotations. However, constructing such labels is costly, time-consuming, and often noisy due to reliance on external per- ception modules. In this work, we propose a novel paradigm inspired by human cognitive development: learning foundational spatial skills through structured block-manipulation tasks. We introduce SpatialBlock-15k, a synthetic dataset of 15,000 block-stacking problems covering 3D-to-2D projection, viewpoint transformation, and structural combination. The dataset further incorporates controlled color modulation as visual cues to encourage anchor-based reasoning in visually complex conditions. Ex- periments demonstrate that LVLMs trained on our dataset through ei- ther direct answering or reasoning-based prediction significantly outperform baselines and generalize to real-world spatial tasks, despite the datasetās synthetic and compact nature. Code and data are available at https: //github.com/rsoohyun/SpatialBlock. 1 Introduction Large VisionāLanguage Models (LVLMs) (Wang et al., 2024; Bai et al., 2025a; Zhu et al., 2025) have demonstrated remarkable performance across a wide range of 2D image under- standing and reasoning tasks. Despite this progress, their ability to mentally reconstruct the 3D structure of a scene from 2D images ā namely, spatial intelligence ā remains lim- ited. This shortcoming poses a fundamental bottleneck for deploying LVLMs in real-world applications that require robust spatial reasoning, such as autonomous driving (Tian et al., 2025) and robotics (Feng et al., 2025). To address this limitation, prior work has extended model architectures by adding a spe- cialized module for learning 3D features (Cheng et al., 2024; Wu et al., 2025) or designed tasks that directly require spatial reasoning (Yang et al., 2025; Yin et al., 2025) to encourage models to learn such capabilities. However, these approaches typically rely on real-world scene question answering datasets that require dense geometric and object-level annotations of the visual context. Constructing such dense labels is not only time-consuming and costly, but also noisy due to the frequent reliance on external modules (e.g., segmentation or depth estimation models), thereby limiting their reliability and scalability. In contrast to methods that depend on annotation-heavy real-scene data, we pursue a more efficient alternative inspired by the way humans develop spatial intelligence. Developmental ā Equal contribution, ā Corresponding Author. 1 arXiv:2609.07064v1 [cs.CV] 7 Sep 2026 Q. Here are images of a 3D structure made of block and the direction of views. Based on the images, how does the 3D structure appear when viewed from the front? ABD To determine front view of 3D structure, we need to imagine looking at it from "front" direction as indicated in second image.This means we are looking at the structure from the bottom-left towards the top-right in first image. ... [explanationomitted] Let's define the columns from left to right (C1, C2, C3) and rows from bottom to top (R1, R2, R3, R4). ... [explanationomitted] Now, let's combine these into a 4x3 grid (Height x Width), representing the front view: (R4) [Yellow] [ ] [ ] ; (R3) [Yellow] [ Red ] [ ] ; (R2 -2nd row) [Yellow] [ Red ] [ Red ] ; (R1) [Yellow] [Yellow] [Yellow] Let's compare this derived view with the given options: ... [option analysisomitted] Therefore, Option B is the correct representation of the 3D structure when viewed from the front. The final answer is $ B$ C 30.0% 36.7% 95.0% Qwen2.5 -7B Gemini- 2.5-flash GPT-5Human 20.8% Figure 1: Performance Gap Between Humans and LVLMs in Block-Stacking Problem. The top row shows a block-stacking task that requires predicting a 2D projec- tion from a given 3D structure. While humans achieve near-perfect accuracy, state-of-the-art LVLMs struggle even with basic spatial reasoning problems. Leading proprietary models exhibit flawed spatial logic, often producing incorrect interpretations of the structure. psychology (Levine et al., 2012; Jirout & Newcombe, 2015) suggests that foundational spa- tial abilities are often cultivated through structured manipulative tools. Specifically, block play has been widely shown to enhance essential components of spatial cognition, including spatial integration and mental simulation (Caldera et al., 1999; Casey et al., 2008). Despite its simplicity, state-of-the-art LVLMs struggle with such problems, suggesting that they lack even fundamental levels of spatial intelligence (Fig. 1). Motivated by this observation, we propose SpatialBlock-15k, a novel synthetic dataset of block-stacking problems designed to improve spatial intelligence. The dataset comprises three categories of spatial reasoning tasks: (1) 3D-to-2D projection, (2) viewpoint transfor- mation, and (3) structural combination. To better approximate real-world spatial reasoning scenarios, where humans interpret scenes by anchoring on a specific object and reason- ing relationships relative to it, we introduce color modulation as a visual cue, encouraging models to identify task-relevant elements in visually complex scenes. Unlike costly real- scene datasets, our dataset can be generated synthetically, enabling clean, scalable, and cost-effective data construction. Experimental results show that LVLMs trained on SpatialBlock-15k, with either direct an- swer prediction or reasoning-based prediction, significantly outperform baseline models, even though the training data are entirely synthetic and relatively small in scale. These find- ings suggest that targeted training on fundamental spatial reasoning tasks can effectively enhance the spatial intelligence of LVLMs and provide a scalable alternative to annotation- heavy real-scene supervision. In summary, our contribution is three-fold: ā¢ We introduce a new perspective on enhancing spatial intelligence in LVLMs by shifting from annotation-heavy real-scene supervision to foundational spatial skill learning through structured synthetic tasks. ā¢ We present SpatialBlock-15k, a scalable synthetic dataset of block-stacking prob- lems spanning 3D-to-2D projection, viewpoint transformation, and structural com- bination, enhanced with controlled color cues for task-relevant reasoning. ā¢ We show that training LVLMs on SpatialBlock-15k significantly improves spatial reasoning performance and generalizes effectively to real-world scenes, despite rely- ing solely on synthetic data. 2 (a) Spatial Composition(b) Mental Simulation(c) Spatial Integration Figure 2: Spatial Abilities Underlying Block-Stacking Reasoning. Solving block- stacking tasks involves (a) inferring latent 3D composition, (b) reasoning under transforma- tion, and (c) integrating multiple components into a coherent structure. 2 Related Work 2.1 Visual Spatial Reasoning in LVLMs Recent efforts to improve spatial intelligence in LVLMs focus on building a dedicated encoder or proposing a specialized training strategy. On the architectural side, some approaches modify the model by introducing spatial tokens into the vision encoder (Tong et al., 2024; Lou et al., 2025), incorporating depth-aware plugin modules (Cheng et al., 2024), or adding spatial encoders to learn 3D features (Wu et al., 2025). Other research focuses on the training process, designing hierarchical schemes that transition from spatial perception to complex reasoning (Ma et al., 2025a; Li et al., 2026; Liu et al., 2026) or encouraging the model to output cognitive maps that encode object position and orientation (Yang et al., 2025; Yin et al., 2025). Despite their effectiveness, these approaches generally require real- scene annotations or pseudo-3D signals from external modules, which limits their scalability and robustness. 2.2 Training Datasets for Spatial Intelligence Various training datasets have been proposed to enhance the spatial intelligence of LVLMs, ranging from simple spatial relations to complex multi-step reasoning.Early bench- marks (Chen et al., 2024; Ma et al., 2025b) leverage external models (e.g., object detection, segmentation, depth, or pose estimation) to extract 3D information and generate low-level QA pairs such as distance, orientation, and spatial relations. To foster higher-level logic, re- cent works (Yang et al., 2025; Ouyang et al., 2025) use 3D annotated video data to construct relative relation and direction reasoning, or more sophisticated tasks such as route planning or spatio-temporal appearance ordering through human annotation. MindCube (Yin et al., 2025) focuses on spatial mental modeling, which requires synthesizing a full scene from par- tial views or inferring arrangements under perspective shifts. Despite these advancements, a significant bottleneck remains: these approaches rely on dense, real-scene 3D annotations, which are labor-intensive and inherently noisy, necessitating costly human annotation. 3 SpatialBlock-15k We introduce SpatialBlock-15k, a scalable synthetic dataset designed to systematically probe and enhance foundational spatial abilities, moving away from the conventional annotation-intensive paradigm that relies on dense real-scene 3D labels. Specifically, we utilize block-stacking problems, which play a pivotal role in the early stages of human spatial cognitive development (Sec.3.1). Furthermore, we incorporate controlled color variations as visual cues to better reflect real-world spatial reasoning conditions (Sec. 3.2). 3.1 Block-Stacking Problems for Spatial Intelligence Activities involving block-stacking require structural spatial understanding that extends beyond superficial visual pattern recognition. As illustrated in Fig. 2, solving such problems requires (a) spatial composition, the ability to reconstruct the complete 3D structure by deconstructing the visible configuration and inferring occluded blocks that must exist to 3 (a) Block-Stacking Problems(b) Visual Cue-based Extension Q.How does 3D structure appear when viewed from the arrow direction? Q.How does 3D structure appear when rotated 270 degrees clockwise? Q. If we push 3D structure in the direction indicated by the arrow, which 3D structure can be formed? Q1 Structure1Structure2 ABCD ABCD ABCD Q2 Q3 Q.How does 3D structure appear when viewed from the arrow direction? Q.How does 3D structure appear when rotated 270 degrees clockwise? Q. If we move Structure 2 so that its yellow and green blocks overlap with the matching ones in Structure 1, which 3D structure can be formed? ABCD ABCD ABCD SpatialBlock-15k Figure 3: SpatialBlock-15k. (a) Three block-stacking tasks: 3D-to-2D projection (Q1), viewpoint transformation (Q2), and structural combination (Q3). (b) Visual cue-based ex- tension with controlled color variations, where colors encode depth ordering (Q1), preserved structural roles across transformations (Q2), and correspondence cues for integration (Q3). support it; (b) mental simulation, the capacity to simulate the outcome of transformations applied to a 3D structure; and (c) spatial integration, the ability to anticipate the resulting structure when multiple components are combined. Together, these abilities reflect core dimensions of spatial intelligence. Building on these three abilities, we define three question types, each targeting a distinct aspect of spatial reasoning (Fig. 3(a)): Q1: 3D-to-2D projection This task requires predicting the 2D appearance of a 3D structure from a specific viewing direction. Solving it demands the ability to in- ternally reconstruct the 3D configuration and mentally transform it to the target viewpoint. The model must additionally perform depth-aware occlusion reasoning to determine which blocks remain visible in the final 2D projection. Q2: Viewpoint transformation The problem inquires how a given 3D structure ap- pears under self-rotation or viewpoint changes. The model should preserve struc- tural consistency across transformations, maintaining correct relative positions of all components without distortion, disappearance, or unintended displacement. Q3: Structural combination This question asks how the resultant integration is formed when two 3D structures are combined. It requires simulating how individual components interact, identifying contact interfaces, and subsequently inferring the coherent global structure that emerges from their integration. Collectively, these question types form the foundation of our SpatialBlock-15k dataset, pro- viding a structured framework for evaluating and training spatial reasoning in LVLMs. 3.2 Visual Cue-based Extension In the previous section, we constructed tasks that require understanding 3D structures composed of single-color blocks. Since color does not provide a discriminative feature, these tasks demand that the model focus on purely geometric relationships, such as the relative positions of blocks and appearance changes under viewpoint transformations. However, real-world scenes typically consist of complex environments with diverse objects, rather than visually uniform ones. Given this intricacy, humans tend to analyze the global 4 scene configurations by selecting a specific object as an anchor and interpreting relationships relative to it (Land & Hayhoe, 2001). That is, visually salient cues serve as the reference points to infer changes and relations. Motivated by this cognitive characteristic, we extend the monochromatic block-based tasks introduced in Sec.3.1 by incorporating color as an additional cue. Here, color is not used merely for visual diversity but as a functional guidance that reveals depth informa- tion, structural correspondences, and contact regions. This design encourages the model to reconstruct the global structure and track transformations by using specific blocks as reference points. Specifically, we extend the three question types as follows (Fig. 3(b)): Q1: 3D-to-2D projection with depth cues To facilitate depth-aware reasoning, dif- ferent colors are assigned based on depth from a given viewpoint. Thus, the task requires not only matching the silhouette but also tracking frontāback relationships using color information. Color functions as an explicit indicator of depth ordering and is designed to clearly reflect the hierarchical organization of the 3D structure. Q2: Viewpoint transformation with anchor blocks Among the blocks composing the 3D structure, one anchor block is assigned a distinct color. This block must maintain the same structural role before and after transformation. Accordingly, the task requires not only recognizing geometric structure but also understanding anchor-relative relationships that are preserved through transformation. Q3: Structural combination with overlapping blocks The attachment location between two structures is represented not by arrows but by the overlap of colored blocks indicated by transparent regions. Each structure is presented as a separate image, so aligning the positions of same-colored blocks across images is needed to predict the combined structure. Color thus acts as a shared reference point linking two independent inputs, inducing reasoning inter-structural integration. This visual cueābased extension is designed to maintain the requirement for understanding the 3D structure itself while additionally enabling the learning of scene organization and relational reasoning centered around anchor objects, as required in realistic visual contexts. Consequently, it leads to meaningful improvement in real-scene spatial reasoning perfor- mance, as empirically demonstrated in Sec. 5.3.2. Finally, the SpatialBlock-15k dataset comprises three question types with visual cueābased extensions, each containing 5k sam- ples. Further details on data construction can be found in Sec. A. 4 Method To foster the spatial intelligence of LVLMs using SpatialBlock-15k, we propose two training strategies, targeting different aspects of the modelās capabilities: (1) direct answer predic- tion, which focuses on cultivating rapid inference by mapping visual inputs directly to their corresponding answers (Sec.4.1), and (2) reasoning-based prediction, which encourages the model to explain its intermediate logical paths that lead to a final answer (Sec. 4.2). 4.1 Direct Answer Prediction Model To maximize the modelās capacity for instantaneous spatial problem-solving, we train it with supervision only on the ground-truth answer sequence y =y 1 ,y 2 ,...,y T . Specifically, given a textual query q and image v, we optimize model parameters Īø to predict the next token y i , given the preceding context y (1:iā1) , by minimizing standard cross-entropy loss: L ce (Īø) =ā X i logP y i | y (1:iā1) ,q,v . (1) This strategy directly optimizes answer precision, thereby establishing a foundation for high-fidelity spatial problem-solving proficiency. 5 4.2 Reasoning-based Prediction Model To go beyond simple answer prediction, we adopt a second strategy that enables the model to articulate its internal logic. To ensure a solid foundation for spatial intelligence, we first initialize the model through supervised fine-tuning to establish basic task-solving ability, then employ reinforcement learning to facilitate high-order structural reasoning. For model initialization, we train the model with Low-Rank Adaptation (LoRA) instead of full-parameter fine-tuning, as full fine-tuning often degrades the inherent reasoning abil- ity. This lightweight adaptation maintains the modelās Chain-of-Thought (CoT) reasoning capabilities while aligning it with our dataset. Notably, we deviate from the conventional ācold-startā phase that relies on synthesized trajectories from larger teacher models (Guo et al., 2025; Li et al., 2026). The rationale behind this decision is that even state-of-the-art proprietary models exhibit relatively low accuracy on our dataset (Tab. 1), making it in- feasible to construct reliable high-quality trajectories. Our empirical analysis (Sec. 5.3.4) demonstrates a clear performance advantage in preserving the modelās reasoning ability through LoRA-based tuning. This approach provides a more effective initialization for RL training than attempting to recover the thinking ability via a low-quality cold-start phase. After LoRA-based initialization, we directly optimize the model using reinforcement learning via Group Relative Policy Optimization (GRPO). We design a multi-objective reward func- tion that evaluates response correctness and reasoning trace quality with three components: R(y,o) =R acc (y,o) +R format (o) +R len (o),(2) where o is a model-generated prediction given a question c = (q,v) with ground-truth answer y. The accuracy reward R acc verifies whether the final answer extracted from o matches y. Next, the format rewardR format enforces the required CoT structure, with a comprehensive reasoning trace followed by the final answer enclosed within <answer> and </answer> tags. Lastly, to prevent overly brief or excessively verbose output, a length reward R len requires the response length L to satisfy 50 < L < 1024. For simplicity and stability, all components are implemented in binary rewards (0 or 1) based on whether each condition is satisfied. Given this reward, for each question c, the old policy model Ļ Īø old samples a group of can- didate responses o 1 ,o 2 ,...,o G . Each response o i receives a reward r i = R(y,o i ), from which we compute the advantage A i = r i āmean(r 1 ,r 2 ,...,r G ) std(r 1 ,r 2 ,...,r G ) . The model is then updated by maximizing the following objective: J GRPO (Īø)=E c,o i h 1 G P G i=1 min Ļ Īø (o i |c) Ļ Īø old (o i |c) A i ,clip Ļ Īø (o i |c) Ļ Īø old (o i |c) ,1Ā±Īµ A i āĪ²KL[Ļ Īø ||Ļ ref ] i , (3) where Īµ and Ī² are hyper-parameters, and KL[Ļ Īø ||Ļ ref ] is the KL divergence between the policy model Ļ Īø and reference model Ļ ref . 5 Experiments We conduct multiple experiments to validate the effectiveness of SpatialBlock-15k in foster- ing spatial intelligence. We first describe the experimental setup, including training details and evaluation protocols (Sec.5.1). We then compare our models with prior methods across several spatial reasoning benchmarks (Sec. 5.2). Next, we perform ablation studies to assess the effects of each component in dataset construction and training (Sec.5.3). Lastly, we provide an analysis of the generalizability of our proposed dataset (Sec. 5.4). 5.1 Experimental Settings We adopt Qwen2.5-VL-3B (Bai et al., 2025b), Qwen2.5-VL-7B (Bai et al., 2025b), Qwen3- VL-4B (Bai et al., 2025a), and InternVL3-2B (Zhu et al., 2025) as baseline models for our framework. To distinguish the training strategy, we append a suffix to the model name: SpatialBlock-direct denotes the model that outputs the answer directly (Sec. 4.1), while SpatialBlock-reason refers to the model that performs reasoning before producing the final answer (Sec. 4.2). Both variants are trained using the entire SpatialBlock-15k dataset. 6 Table 1: Main Results. For each dataset, bold and underlinedindicate the best and second-best performance, respectively. * represents in-domain results where the model was trained on the benchmark. We re-evaluated all previous works for fair comparison. Model# Data In-domainOut-of-domain SB-BenchMindCubeMMSI-BenchSPBenchMMMUOverall Proprietary Models GPT-5 (Singh et al., 2025)-43.856.742.843.077.655.0 Gemini-2.5-flash (Comanici et al., 2025)-36.451.032.244.358.846.6 Claude-Sonnet-4.5 (Anthropic, 2025)-46.334.032.743.969.243.9 Open-Source Models LLaVA-OneVision-7B (Li et al., 2024)-23.839.028.443.442.639.2 InternVL3-8B (Zhu et al., 2025)-22.742.725.146.148.040.5 Llama-3.2-11B-Vision-Instruct (Dubey et al., 2024)-23.029.524.528.240.730.7 Spatial Specialist Models Qwen2.5-VL-3B Based Models Qwen2.5-VL-3B-Instruct (Bai et al., 2025b)-29.939.126.140.045.637.7 SpaceQwen (Chen et al., 2024)2B30.638.824.540.8 45.637.4 SpatialMLLM (Wu et al., 2025)120K30.036.325.440.840.435.7 SpatialLadder-3B (Li et al., 2026)25K25.546.426.076.6*47.339.9 Spatial-SSRL-3B (Liu et al., 2026)81K29.541.024.419.545.232.5 SpatialBlock-3B-direct (Ours)15K94.849.128.140.446.441.0 SpatialBlock-3B-reason (Ours)15K90.245.329.243.746.741.2 Qwen2.5-VL-7B Based Models Qwen2.5-VL-7B-Instruct (Bai et al., 2025b)-20.831.327.446.746.6 38.0 SpaceR (Ouyang et al., 2025)150K26.831.627.453.155.441.9 Spatial-SSRL-7B (Liu et al., 2026)81K27.735.127.523.448.933.7 SpatialBlock-7B-direct (Ours)15K95.048.828.448.745.242.8 SpatialBlock-7B-reason (Ours)15K77.537.929.750.955.543.5 Qwen3-VL-4B Based Models Qwen3-VL-4B-Instruct (Bai et al., 2025a)-17.026.227.144.547.536.3 Spatial-SSRL-4B (Liu et al., 2026)81K34.834.628.323.150.534.1 SpatialBlock-4B-direct (Ours)15K95.751.327.845.947.243.1 SpatialBlock-4B-reason (Ours)15K94.5 41.128.343.749.640.7 InternVL3-2B Based Models InternVL3-2B-Instruct (Zhu et al., 2025)-19.032.127.033.041.333.4 SpatialBlock-2B-direct (Ours)15K97.251.427.341.342.440.6 SpatialBlock-2B-reason (Ours)15K78.338.927.845.647.339.9 To evaluate the spatial reasoning capability, we conduct experiments on five benchmarks spanning both in-domain and out-of-domain settings. For in-domain evaluation, we intro- duce SB-Bench, the held-out test split of SpatialBlock-15k, comprising 600 questions. For out-of-domain evaluation, we assess generalization to real-world scenarios on four bench- marks: MindCube (Yin et al., 2025), MMSI-Bench (Yang et al., 2026), and SPBench (Li et al., 2026) for spatial reasoning, and MMMU (Yue et al., 2024) for general visual per- ception. Across all benchmarks, we restrict evaluation to multiple-choice questions to align with our training answer format. Additional numerical results are provided in Sec. C.1. 5.2 Main Results Tab. 1 shows our methodās performance on spatial reasoning benchmarks. Notably, despite being trained with only 15k synthetic SpatialBlock samples, our model consistently outper- forms existing spatial specialists on real-scene benchmarks. This result highlights that our block-stacking formulation provides an effective training signal for spatial reasoning. For SpatialBlock-direct models, we observe substantial gains, particularly on MindCube, which requires mental simulation. Among Qwen-based models, the 3B model outperforms the state-of-the-art SpatialLadder by 2.7%, while the 7B model shows an even larger gain of 17.6% over its backbone. The 4B model further improves upon its backbone by 25.1%, achieving 51.3%, the best open-source performance. Beyond Qwen, our method also im- proves InternVL3, demonstrating its generalizability across model scales and families. The reasoning-enhanced variant, SpatialBlock-reason, further demonstrates strong re- sults on MMSI-Bench, which requires complex logical inference. The 3B model outperforms SpatialLadder-3B by 3.2% using a simpler training pipeline based solely on synthetic block- stacking data, whereas the 7B model outperforms SpaceR and Spatial-SSRL with only 15K training samples. Beyond final-answer accuracy, our model also improves reasoning quality, achieving a higher alignment score of 21.1 with ground-truth reasoning steps, compared to 17.8 for the baseline. (See Sec. B.4 for details.) On SPBench, our models consistently 7 Table 2: Ablation Studies. * denotes in-domain results from benchmark-trained models. Model# Data In-domainOut-of-domain SB-BenchMindCubeMMSI-BenchSPBenchMMMUOverall Qwen2.5-VL-3B-Instruct (Bai et al., 2025b)-29.939.126.140.045.637.7 SpatialBlock-3B-direct (Ours)15K94.849.128.140.446.4 41.0 SpatialBlock-3B-reason (Ours)15K90.245.329.243.746.741.2 (a) Ablation on question types SpatialBlock-3B-reason (Q1 only)15K48.041.827.642.843.238.8 SpatialBlock-3B-reason (Q2 only)15K51.538.824.242.147.138.0 SpatialBlock-3B-reason (Q3 only)15K40.841.325.343.246.539.1 (b) Ablation on visual cues SpatialBlock-3B-direct (w/o visual cues)15K61.741.226.238.446.638.1 SpatialBlock-3B-reason (w/o visual cues)15K55.838.024.444.848.839.0 (c) Ablation on task design SpatialBlock-3B-direct (SpatialLadder-26k)15K24.744.527.272.1*45.347.3 SpatialBlock-3B-reason (SpatialLadder-26k)15K28.044.927.568.8*44.7 46.5 (d) Ablation on model initialization SpatialBlock-3B-reason (No Init.)15K67.841.325.443.546.739.2 SpatialBlock-3B-reason (Full-SFT Init., Qwen2.5-VL-7B)15K96.543.027.647.844.7 40.8 SpatialBlock-3B-reason (Full-SFT Init., GPT-5)15K75.341.728.044.348.1 40.5 improve relative spatial reasoning despite such questions not being explicitly included in the training data, demonstrating strong generalization beyond the supervised task distribution. In summary, direct models excel on benchmarks that require canonical 90-degree view- point transformations, while reasoning models perform better on tasks requiring diverse viewpoint changes and multi-image reasoning. Also, despite being trained without any real- scene images, our models maintain stable performance on the general visual understanding benchmark MMMU, indicating that SpatialBlock training enhances 3D spatial reasoning without degrading broader visual perception. 5.3 Ablation Studies 5.3.1 Ablation on Question Types We evaluate models trained on a single question type (Tab.2(a)). Each question type targets a distinct aspect of spatial reasoning: Q1 focuses on viewpoint changes, Q2 em- phasizes spatial transformations such as 90-degree rotations, and Q3 requires integrating spatial information across multiple perspectives. The single-type variants exhibit different performance patterns across benchmarks, suggesting that each type provides distinct spa- tial supervision. However, while single-type variants show limited gains, training on all three types achieves the best performance across most benchmarks. These results indicate that the three types provide complementary supervision, and combining them leads to more robust spatial reasoning and generalization. 5.3.2 Ablation on Visual Cues We introduce color manipulation as a visual cue to approximate how humans interpret real- world scenes using visually salient reference objects (Sec. 3.2). To evaluate its effectiveness, we compare models trained with and without visual cues. As shown in Tab. 2(b), removing visual cues leads to performance drops in direct and reason models of 7.9% and 7.3% in MindCube, and 1.9% and 4.8% in MMSI-Bench, respectively. These results suggest that color cues help models better understand complex scenes and mimic human scene perception, improving spatial reasoning in real-world environments. 5.3.3 Ablation on Task Design In order to analyze the effect of the task design itself more thoroughly, we train a model on SpatialLadder-26k using the identical training recipe. For a fair comparison, we use 11.5k multiple-choice questions matching our response format and train for 1.3 epochs to match the same training size. As shown in Tab. 2(c), this model underperforms ours on most benchmarks except the in-domain benchmark SPBench, suggesting that our performance gains mainly stem from the proposed block-stacking task design. 8 (a) Relative Direction(b) Relative Distance From the cameraās perspective, where isthe clock relative to keyboard? A. left-front B. left-back C. right-frontD. right-back Real-scene Based on the image, where is the red block located relative to the yellow block? A. front B. left-behind C. left-front D. behind Synthetic Which of these two objects (radiator, chair) is closer to the trash can? A. radiator B. chair Based on the image, which color block is closer to the yellow block? A. purple B. green Real-scene Synthetic Figure 4: Synthetic-Real Dataset. Widely used spatial question types, (a) relative direc- tion and (b) relative distance, are adapted into a block-based environment while preserving their semantic structure. Table 3: Comparison of Training on SpatialBlock and Synthetic-Real. While Synthetic-Real improves in-domain performance, training on SpatialBlock leads to better generalization on out-of-domain benchmark datasets. All experiments are conducted on Qwen2.5-VL-3B (Bai et al., 2025b), and in-domain performance is evaluated on SPBench. In-domainOut-of-domain Train DatasetRelative DirectionRelative DistanceMindCubeMMSI-BenchMMMU Baseline31.474.139.126.145.6 Synthetic-Real43.476.932.825.543.6 SpatialBlock31.467.449.128.146.4 5.3.4 Ablation on model initialization for reasoning model To train a reasoning-capable model via reinforcement learning, we adopt a LoRA-based initialization strategy to enhance baseline spatial intelligence while preserving inherent rea- soning ability. Tab. 2(d) compares different initialization strategies and their effects on performance. Without prior fine-tuning, performance drops noticeably across benchmarks, indicating the importance of establishing basic spatial understanding. We further compare our approach with commonly used cold-start phases based on full-parameter tuning (Guo et al., 2025; Li et al., 2026). While cold-start training can improve instruction following and achieve higher accuracy convergence, generating reliable trajectories for our tasks is challenging: both a larger model (Qwen2.5-VL-7B) and a proprietary model (GPT-5) fre- quently produce inaccurate block descriptions and reasoning. Consequently, training with these noisy reasoning traces yields only marginal gains. In contrast, LoRA-based initializa- tion provides substantial improvements, offering a strong starting point for reinforcement learning with both basic spatial intelligence and reliable instruction-following ability. 5.4 Generalizability of Block-Stacking Questions As shown in Tab. 1, our models trained on SpatialBlock-15k generalize well to real-scene benchmark datasets despite involving question types that differ from those commonly used. This observation raises the question of whether our block-stacking tasks provide a more effective learning signal than conventional spatial question formulations. To investigate this, we construct Synthetic-Real, a dataset of the same size as SpatialBlock-15k, which adapts two widely used spatial question types ā relative direction and distance ā into our block- based environment (Fig. 4). Although both datasets can be generated with comparable effort, the results in Tab.3 reveal that the model trained on Synthetic-Real achieves higher accuracy only on in-domain tasks and performs poorly on out-of-domain benchmark datasets. In contrast, our model consistently improves across benchmarks. These results suggest that SpatialBlock tasks promote a more holistic understanding of 3D scenes, leading to more generalizable and transferable spatial reasoning than standard spatial supervision. 9 6 Conclusion In this paper, we revisited the problem of improving spatial intelligence in large vision- language models from a new perspective: learning foundational spatial skills through struc- tured tasks rather than relying on annotation-heavy real-scene supervision. To this end, we introduced SpatialBlock-15k, a novel synthetic dataset of block-stacking problems, includ- ing 3D-to-2D projection, viewpoint transformation, and structural integration. The dataset further incorporated controlled color variation as visual cues to encourage anchor-based reasoning, mimicking how humans interpret scenes by using specific objects as reference points. Using this dataset, we developed two specialized models: SpatialBlock-direct for immediate answer generation and SpatialBlock-reason for detailed logical inference. Experimental results demonstrated that our models consistently outperform existing spatial specialists on various real-world spatial benchmarks. Furthermore, extensive analyses show that the learned capabilities generalize across tasks and domains, indicating that the models capture the underlying spatial structure of scenes. These findings highlighted that robust spatial intelligence can be achieved through training on carefully designed synthetic spatial tasks, offering a scalable and effective path for improving spatial reasoning in LVLMs. References Anthropic.Claude sonnet 4.5 system card.Technical report,Anthropic, 2025.URL https://assets.anthropic.com/m/12f214efcc2f457a/original/ Claude-Sonnet-4-5-System-Card.pdf. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025a. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report, 2025b. URL https://arxiv.org/abs/2502.13923. Yvonne M Caldera, Anne McDonald Culp, Marion OāBrien, Rosemarie T Truglio, Mil- dred Alvarez, and Aletha C Huston. Childrenās play preferences, construction play with blocks, and visual-spatial skills: Are they related? International Journal of Behavioral Development, 23(4):855ā872, 1999. Beth M Casey, Nicole Andrews, Holly Schindler, Joanne E Kersh, Alexandra Samper, and Juanita Copley. The development of spatial skills through interventions involving block building activities. Cognition and instruction, 26(3):269ā309, 2008. Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 14455ā14465, 2024. An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatialrgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062ā135093, 2024. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, In- derjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next gen- eration agentic capabilities. arXiv preprint arXiv:2507.06261, 2025. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony S. Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurāelien Rodriguez, Austen Gregerson, Ava 10 Spataru, Baptiste Rozi`ere, Bethany M. Biron, Binh Tang, Bobbie Chern, Char lotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Niko- laidis, Damien Allonsius, Daniel J. Song, Danielle Pintz, Danny Livshits, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hup- kes, Egor Lakomkin, Ehab A. AlBadawy, E I Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Graeme Nail, Gr Ģegoire Mialon, Guanglong Pang, Guillem Cu-curell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel M. Kloumann, Ishan Misra, Ivan Evtimov, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Ju-Qing Jia, Kalyan Vasuden Alwala, K. Upasani, Kate Plawiak, Keqian Li, Kenneth Heafield, Kevin R. Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuen ley Chiu, Kunal Bhalla, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke de Oliveira, Madeline Muzzi, Ma hesh Pasupuleti, Mannat Singh, Manohar Paluri, Marcin Kardas, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melissa Hall Melanie Kambadur, Mike Lewis, Min Si, Mitesh Kumar Singh, Mona Has- san, Naman Goyal, Narjes Torabi, Nikolay Bash-lykov, Nikolay Bogoychev, Niladri S. Chatterji, Olivier Duchenne, Onur cCelebi, Patrick Alrassy, Pengchuan Zhang, Peng- wei Li, Petar Vasi Ģc, Peter Weng, Prajjwal Bhargava, Pratik Dubal, Praveen Krishnan, Punit Singh Koura, Puxin Xu, Qing He, Qingxiao Dong, R. S. Mughil Srinivasan, Raj Ganapathy, Ramon Calderer, Ricardo Silveira Cabral, Robert Stojnic, Roberta Raileanu, Rohit Girdhar, Rohit Patel, Romain Sauvestre, Ron nie Polidoro, Roshan Sumbaly, Ross Taylor, Ruan Silva, Rui Hou, Rui Wang, Saghar Hosseini, Sa hana Chennabasappa, San- jay Singh, Sean Bell, Seohyun Sonia Kim, Sergey Edunov, Shaoliang Nie, Sharan Narang, Sharath Chandra Raparthy, Sheng Shen, Shengye Wan, Shruti Bhosale, Shun Zhang, Simon Vandenhende, Soumya Batra, Spencer Whit-man, Sten Sootla, St Ģephane Collot, Suchin Gururangan, Sydney Borodinsky, Tamar Herman, Tara Fowler, Tarek Sheasha, Thomas Georgiou, Thomas Scialom, Tobias Speckbacher, Todor Mihaylov, Tong Xiao, Ujjwal Karn, Vedanuj Goswami, Vibhor Gupta, Vignesh Ramanathan, Viktor Kerkez, Vincent Gonguet, Virginie Do, Vish Vogeti, Vladan Petrovic, Weiwei Chu, Wenhan Xiong, Wenyin Fu, Whit ney Meers, Xavier Martinet, Xiaodong Wang, Xiaoqing Ellen Tan, Xin- feng Xie, Xuchao Jia, Xuewei Wang, Yaelle Goldschlag, Yashesh Gaur, Yasmine Babaei, Yiqian Wen, Yiwen Song, Yuchen Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coud- ert, Zhengxu Yan, Zhengxing Chen, Zoe Papakipos, Aaditya K. Singh, Aaron Grattafiori, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adi Gangidi, Adolfo Victoria, Ahuva Gold- stand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alex Vaughan, Alexei Baevski, Allie Fein-stein, Amanda Kallet, Amit Sangani, Anam Yunus, Andrei Lupu, Andres Alvarado, Andrew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ram- chandani, Annie Franco, Aparajita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, Ash- win R. Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leonhardi, Po-Yao (Bernie) Huang, Beth Loyd, Beto de Paola, Bhargavi Paranjape, Bing Liu, Bo Wu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Montalvo, Carl Parker, Carly Burton, Catalina Mejia, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, Ching-Hsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Damon Civin, Dana Beaty, Daniel Kreymer, Shang- Wen Li, Danny Wyatt, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Ed- ward Dowling, Eissa Jamil, Elaine Montgomery, Eleonora Presani, Emily Hahn, Emily Wood, Erik Brinkman, Esteban Arcaute, Evan Dunbar, Evan Smoth-ers, Fei Sun, Felix Kreuk, Feng Tian, Firat Ozgenel, Francesco Caggioni, Francisco (Paco) Guzm Ģan, Frank J. Kanayet, Frank Seide, Gabriela Medina Florez, Gabriella Schwarz, Gada Badeer, Geor- gia Swee, Gil Halpern, Govind Thattai, Grant Herman, Grigory Sizov, Guangyi Zhang, Guna Lakshminarayanan, Hamid Shojanazeri, Han Zou, Hannah Wang, Han Zha, Haroun Habeeb, Harrison Rudolph, Helen Suk, Henry Aspegren, Hunter Goldman, Igor Molybog, Igor Tufanov, Irina-Elena Veliche, Itai Gat, Jake Weissman, James Geboski, James Kohli, 11 Japhet Asher, Jean-Baptiste Gaya, Jeff Marcus, Jeff Tang, Jennifer Chan, Jenny Zhen, Jeremy Reizenstein, Jeremy Teboul, Jessica Zhong, Jian Jin, Jingyi Yang, Joe Cum- mings, Jon Carvill, Jon Shepard, Jonathan McPhie, Jonathan Torres, Josh Ginsburg, Junjie Wang, Kaixing(Kai) Wu, U KamHou, Karan Saxena, Karthik Prasad, Kartikay Khandelwal, Katayoun Zand, Kathy Matosich, Kaushik Veeraraghavan, Kelly Michelena, Keqian Li, Kun Huang, Kunal Chawla, Kushal Lakhotia, Kyle Huang, Lailin Chen, Lak- shya Garg, A Lavender, Leandro Silva, Lee Bell, Lei Zhang, Liangpeng Guo, Licheng Yu, Liron Moshkovich, Luca Wehrstedt, Madian Khabsa, Manav Avalani, Manish Bhatt, Maria Tsimpoukelli, Martynas Mankus, Matan Hasson, Matthias Lennie, Matthias Reso, Maxim Groshev, Maxim Naumov, Maya Lathi, Meghan Keneally, Michael L. Seltzer, Michal Valko, Michelle Re-strepo, Mihir Patel, Mik Vyatskov, Mikayel Samvelyan, Mike Clark, Mike Macey, Mike Wang, Miquel Jubert Hermoso, Mo Metanat, Mohammad Raste- gari, Mun ish Bansal, Nandhini Santhanam, Natascha Parks, Natasha White, Navy ata Bawa, Nayan Singhal, Nick Egebo, Nicolas Usunier, Nikolay Pavlovich Laptev, Ning Dong, Ning Zhang, Norman Cheng, Oleg Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pe dro Rittner, Philip Bon- trager, Pierre Roux, Piotr Doll Ģar, Polina Zvyagina, Prashant Ratanchandani, Pritish Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Raghotham Murthy, Raghu Nayani, Rahul Mitra, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Ro han Maheswari, Russ Howes, Ruty Rinott, Sai Jayesh Bondu, Samyak Datta, Sara Chugh, Sara Hunt, Sargun Dhillon, S. Yu. Sidorov, Satadru Pan, Saurabh Verma, Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Zha, Shiva Shankar, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudar- shan Govindaprasad, Sumit Kumar Gupta, Sung-Bae Cho, Sunny Virk, Suraj Subrama- nian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Kohler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria O Ajayi, Victoria Montanez, Vijai Mohan, Vinay Kumar, Vishal Mangla, Vlad Ionescu, Vlad Andrei Poenaru, Vlad T. Mihailescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xia Tang, Xiaofang Wang, Xiaojian Wu, Xiaolan Wang, Xide Xia, Xilun Wu, Xinbo Gao, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu Wang, Yuchen Hao, Yundi Qian, Yuzi He, Zach Rait, Zachary DeVito, Zef Rosnbrick, Zhaoduo Wen, Zhenyu Yang, and Zhiwei Zhao. The llama 3 herd of models, 2024. URL https://arxiv.org/abs/2407.21783. Zhiyuan Feng, Zhaolu Kang, Qijie Wang, Zhiying Du, Jiongrui Yan, Shubin Shi, Chengbo Yuan, Huizhi Liang, Yu Deng, Qixiu Li, et al. Seeing across views: Benchmarking spatial reasoning of vision-language models in robotic scenes. arXiv preprint arXiv:2510.19400, 2025. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. Jamie J Jirout and Nora S Newcombe. Building blocks for developing spatial skills: Evidence from a large, representative us sample. Psychological science, 26(3):302ā310, 2015. Michael F Land and Mary Hayhoe. In what ways do eye movements contribute to everyday activities? Vision research, 41(25-26):3559ā3565, 2001. Susan C Levine, Kristin R Ratliff, Janellen Huttenlocher, and Joanna Cannon. Early puzzle play: a predictor of preschoolersā spatial transformation skill. Developmental psychology, 48(2):530, 2012. Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. Hongxing Li, Dingming Li, Zixuan Wang, Yuchen Yan, Hang Wu, Wenqi Zhang, Yongliang Shen, Weiming Lu, Jun Xiao, and Yueting Zhuang. Spatialladder: Progressive training 12 for spatial reasoning in vision-language models. In International Conference on Learning Representations, volume 2026, p. 76566ā76592, 2026. Yuhong Liu, Beichen Zhang, Yuhang Zang, Yuhang Cao, Long Xing, Xiaoyi Dong, Haodong Duan, Dahua Lin, and Jiaqi Wang. Spatial-ssrl: Enhancing spatial understanding via self-supervised reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 9570ā9581, 2026. Haoran Lou, Chunxiao Fan, Ziyan Liu, Yuexin Wu, and Xinliang Wang. Llava-sp: En- hancing visual representation with visual spatial tokens for mllms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 22014ā22024, 2025. Wufei Ma, Yu-Cheng Chou, Qihao Liu, Xingrui Wang, Celso de Melo, Jianwen Xie, and Alan Yuille. Spatialreasoner: Towards explicit and generalizable 3d spatial reasoning. arXiv preprint arXiv:2504.20024, 2025a. Wufei Ma, Luoxin Ye, Celso M de Melo, Alan Yuille, and Jieneng Chen. Spatialllm: A com- pound 3d-informed design towards spatially-intelligent large multimodal models. In Pro- ceedings of the Computer Vision and Pattern Recognition Conference, p. 17249ā17260, 2025b. Kun Ouyang, Yuanxin Liu, Haoning Wu, Yi Liu, Hao Zhou, Jie Zhou, Fandong Meng, and Xu Sun. Spacer: Reinforcing mllms in video spatial reasoning. arXiv preprint arXiv:2504.01805, 2025. Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. Openai gpt-5 system card. arXiv preprint arXiv:2601.03267, 2025. Kexin Tian, Jingrui Mao, Yunlong Zhang, Jiwan Jiang, Yang Zhou, and Zhengzhong Tu. Nuscenes-spatialqa: A spatial understanding and reasoning benchmark for vision- language models in autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 4567ā4576, 2025. Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri Iyer, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310ā87356, 2024. Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xue- jing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language modelās perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. Diankun Wu, Fangfu Liu, Yi-Hsin Hung, and Yueqi Duan. Spatial-mllm: Boosting mllm capabilities in visual-based spatial intelligence. arXiv preprint arXiv:2505.23747, 2025. Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in space: How multimodal large language models see, remember, and recall spaces. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 10632ā10643, 2025. Sihan Yang, Runsen Xu, Yiman Xie, Sizhe Yang, Mo Li, Jingli Lin, Chenming Zhu, Xiaochen Chen, Haodong Duan, Xiangyu Yue, et al. Mmsi-bench: A benchmark for multi-image spatial intelligence. In International Conference on Learning Representations, volume 2026, p. 157051ā157088, 2026. Junting Lu Yaowei Zheng, Zhangchi Feng Shenzhi Wang, Yuwen Xiong Dongdong Kuang, and Richong Zhang. Easyr1: An efficient, scalable, multi-modality rl training framework. https://github.com/hiyouga/EasyR1, 2025. Baiqiao Yin, Qineng Wang, Pingyue Zhang, Jianshu Zhang, Kangrui Wang, Zihan Wang, Jieyu Zhang, Keshigeyan Chandrasegaran, Han Liu, Ranjay Krishna, et al. Spatial mental modeling from limited views. In Structural Priors for Vision Workshop at ICCVā25, 2025. 13 Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 9556ā9567, 2024. Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025. 14 A Details on Construction and Evaluation of SpatialBlock-15k A.1 Data Construction Process All 3D structures in SpatialBlock-15k are constructed on a 3Ć3 spatial grid, where each grid position can contain between one and four vertically stacked blocks. The height of each stack defines the underlying 3D configuration. To avoid ambiguity when interpreting rendered structures, we restrict the set of valid configurations so that the number of blocks at every grid position can be reliably inferred from the rendered image. In particular, configurations that have severe occlusion ā where blocks are completely hidden behind others from the viewing perspective ā are excluded. This constraint ensures that the global structure remains visually interpretable while still allowing a diverse set of configurations. Based on the resulting valid structures, we generate the three types of tasks illustrated in Fig. 3. For each task, the distractor choices are carefully constructed to be structurally plausible while incorrect, preventing simple heuristics and encouraging models to reason about the spatial relationships within the structure. A.1.1 Q1: 3D-to-2D projection with depth cues In this task, the model must predict the correct 2D projection (e.g., front or side view) of a given 3D block structure. The wrong choices are generated through controlled perturbations of the correct projection: ā¢ Block-removal perturbation. A visible top block from one column is removed, producing a projection that corresponds to a slightly reduced stack height while remaining structurally plausible. ā¢ Color-consistency perturbation. For a column containing multiple colors, the color of one block is replaced with that of a neighboring block in the same column. This preserves the geometric structure but violates the correct color arrangement. ā¢ Combined perturbation. A hybrid modification that simultaneously removes a block from one column and alters the color assignment in another column. During choice generation, perturbations are only applied to columns that contain visible blocks and sufficient color diversity, ensuring that distractors remain meaningful and visually distinguishable. A.1.2 Q2: Viewpoint transformation with anchor blocks This task evaluates whether models can reason about global spatial transformations of the structure. Given an original structure, one transformation is randomly selected from the following set: 90 ā¦ , 180 ā¦ , or 270 ā¦ rotations, and flipping vertically. A specific visible block in the original structure is highlighted, and the model needs to identify the corresponding block position after the transformation. To ensure valid supervision signals, only blocks that remain visible after transformation are selected as colored blocks. The answer choices consist of four candidates: ā¢ the correctly transformed structure with the correct mapped block position (correct answer), ā¢ the original structure with a different block highlighted, ā¢ the correctly transformed structure with an incorrectly highlighted block, ā¢ or a structure generated through a different transformation. A.1.3 Q3: Structural combination with overlapping blocks This task focuses on reasoning about the union when two block structures are combined. Each problem instance consists of two structures and a specified attachment point indicated 15 by overlapping color. The goal is to determine the correct outcome of the combination. During dataset construction, we sample a possible combination from the given 3D structures and compute their first contact location as the overlap region. The distractor choices are constructed by choosing a 3D configuration where one component is replaced, and the other remains unchanged, resulting in a different structure. By ensuring that each distractor differs only in one of the two input structures, the choices remain visually similar to the correct answer while requiring the model to reason about the precise spatial interaction between the specific pair of structures. A.2 Human Evaluation on SB-Bench For human evaluation, we randomly select 30 samples from SB-Bench and collect responses from 31 participants. As illustrated in Fig. 1, humans achieve 95% accuracy on average, demonstrating the taskās simplicity. B Training Implementation Details B.1 Early stopping for SpatialBlock-direct To prevent overfitting to the block structure during direct model training, we selected the model from an intermediate iteration in which test accuracy exceeded 93%, specifically the iteration immediately before it began to decrease. B.2 Reasoning prompt used for reinforcement learning The detailed instructions for reinforcement learning are shown in Fig.5. The prompt guides the model to output reasoning steps in a numbered sequence and gives the final answer between <answer> and </answer> tags. We do not use the extra <reason> tag to reduce the load of learning the format. Please think about this question step by step and provide your detailed reasoning in a numbered reasoning sequence (1., 2., 3., ...). Then give your final answer between the <answer> </answer> tags. Output must be EXACTLY in this format (including tags): ... <answer>X</answer> Please provide only the single option letter (e.g., A, B, C, D, etc.) within the <answer> </answer> tags. RL prompt Figure 5: Instructions for Reinforcement Learning. The model is instructed to provide a step-by-step thinking process, and then the final answer between answer tags. B.3 Hyperparameters used for training We apply LoRA-based supervised fine-tuning for SpatialBlock-reason, using rank 8 for the 3B and 2B models and rank 16 for the 7B and 4B models. For reinforcement learning, we use EasyR1 (Yaowei Zheng et al., 2025) as our base code and follow its default hyperparameters. We set the batch size to 512 with 5 rollouts, while extending the maximum prompt length to 4096 to accommodate multi-image training. All training was conducted for 1 epoch using SpatialBlock-15k. B.4 Reasoning quality evaluation protocol To further assess reasoning quality, we utilize MMSI-Bench with its ground-truth rationales. GPT-5 scores each reasoningās alignment with each ground-truth step as 1 (present), 0.5 (implicit), or 0 (incorrect), and the final score is averaged and scaled to 100. The detailed instructions are shown in Fig. 6. 16 TASK : Compare the reasoning of Model 1 and Model 2, given the Ground Truth (GT) reasoning. First, break the GT reasoning into its core logical step. For each GT reasoning steps, label each model's reasoning as present/implicit/incorrect. Finally output the final winning model between <answer> and </answer>. If they are both equally bad or good, output None. GT reasoning : gt Model 1 : model_1 Model 2 : model_2 Format : <GT> Step1: [reasoning], ... </GT> <Model 1> Step1 : [label], ... </Model 1> <Model 2> Step2 : [label], ... </Model 2> <answer> Model X </answer> GPT-eval prompt Figure 6: Instructions for Reasoning Quality Evaluation. We evaluate the quality of model-generated reasoning by measuring its alignment with each step of the corresponding ground-truth rationale. C Additional Results C.1 Additional Results on Numerical Tasks Table 4: Results on Numerical Tasks. Results on tasks from SPBench. ModelAbsolute distanceObject sizeObject countingAverage Qwen2.5-VL-3B-Instruct30.917.635.127.9 SpatialBlock-3B-direct23.219.972.138.4 SpatialBlock-3B-reason21.438.343.334.3 We evaluate our model on numerical prediction tasks from SPBench (Tab.4). Despite being trained only on multiple-choice questions, our model improves in counting and size estimation, as our dataset enhances the modelās ability to individuate objects and infer their scales. In contrast, it is less effective at absolute distance because our method prioritizes structural understanding and spatial integration over precise metric regression. C.2 Robustness to viewing angle variations Q.How does 3D structure appear when viewed from the arrow direction? 30Ā°-view60Ā°-view ABCD Train dataset (a) Block Structures Under Different Viewing Angles(b) Performance 40 20 0 60 100 80 3B-direct3B-reason 94.894.8 94.2 90.2 89.0 89.5 45Ā°-view Figure 7: Performance Across Viewing Angles. (a) The same block structures are rendered from different viewing angles. (b) Performance of models across viewing angles; the bars correspond to results under 45 ā¦ , 30 ā¦ , and 60 ā¦ views, respectively. To further evaluate whether the model learns the underlying 3D structure rather than mem- orizing appearance patterns from a fixed viewpoint, we conduct a viewpoint generalization experiment. Our SpatialBlock-15k dataset is constructed using blocks rendered from a 45 ā¦ viewing angle, which is fixed during training. We generate the same problems using identi- cal block configurations rendered from different viewing angles (30 ā¦ and 60 ā¦ ) (Fig. 7(a)). As shown in Fig. 7(b), both the direct answering and reasoning-based models maintain consistent performance across these viewpoints. This suggests that the models capture the underlying structure of the blocks, demonstrating genuine enhancement in spatial ability. 17 C.3 Cold-start data quality To determine how the 3D structure appears when viewed from the side, let's analyze the structure step by step: 1. **Identify the structure**: The structure is composed of blocks arranged in a specific pattern. There are two main sections: one on the left and one on the right. 2. **Left section**: The left section has a tall stack of yellow blocks with a smaller stack of blue blocks on top. This stack is aligned vertically. 3. **Right section**: The right section has a larger stack of blue blocks with a smaller stack of yellow blocks on top. This stack is also aligned vertically but is wider than the left section. 4. **Side view**: When viewed from the side, we need to see the vertical alignment of the blocks. The left section will appear as a tall column of yellow blocks with a smaller blue block on top. The right section will appear as a wider column of blue blocks with a smaller yellow block on top . 5. **Compare with options**: - Option A shows a tall column of yellow blocks with a smaller blue block on top, which matches the left section. - Option B shows a wide column of blue blocks with a smaller yellow block on top, which matches the right section. - Option C shows a tall column of yellow blocks with a smaller blue block on top, which matches the left section. - Option D shows a wide column of blue blocks with a smaller yellow block on top, which matches the right section. Since the side view should show both sections, the correct option is the one that combines both the left and right sections. <answer>C</answer> Qwen2.5-VL-7B Q. Here are images of a 3D structure made of block. Based on the images, how does the 3D structure appear when viewed from the arrow direction? ABDC 1. From the right (side) view, heights by depth are: front = 2, middle = 2, back = 3. 2. The nearest layer at the bottom across all depths is blue, so the entire bottom row is blue. 3. At the second level, front and middle are still blue (blue stacks of height 2), while the back shows yellow (taller yellow stack behind ). 4. Only the back column reaches a third level, which is yellow. This matches option C. <answer>C</answer> GPT-5 Figure 8: Generated Cold-Start Data Sample. The reasoning trace extracted from Qwen2.5-VL-7B includes incorrect information about the given 3D structure, even though the model answers correctly. As mentioned in Sec.5.3.4, full-parameter fine-tuning with a cold-start phase shows marginal improvement compared with LoRA-based initialization. We hypothesize that this is due to the low quality of the cold-start reasoning trace. Fig. 8 shows a reasoning ex- ample extracted from Qwen2.5-VL-7B. Although we choose the one with correct answers, the model fails to explain the 3D structure correctly, particularly miscounting the blocks. Furthermore, despite the attempt to evaluate all options, it generates the same explanation for distinct options, leading to a random final choice. We also observed that even proprietary models such as GPT-5 frequently generate inaccurate block descriptions and reasoning (Fig. 8), highlighting the difficulty of producing stable trajectories even for top-tier models on our tasks. Using these GPT-5-generated labels for cold-start SFT still underperforms ours, especially on spatial reasoning benchmarks (Tab. 2(d)), which supports our claim that low-quality teacher reasoning can serve as noisy supervision. 18 C.4 Qualitative Result Q.Based on these three images showing the same scene from different viewpoints (back, left, and right): From the viewpoint presented in image 2, is there a yellow-plastic container behind the grey parking meter? A. YesB. No. Image 1Image 2Image 3 1. First, let's analyze the information provided in image 2: The image shows a grey parking meter on a sidewalk. There is a red sign with white text to the left of the parking meter. 2. Next, let's consider the question: The question asks if there is a yellow-plastic container behind the grey parking meter. Weneed to determine if such a container exists based on the information given in image 2. 3. Now, let's examine the details in image 2: The focus is on the grey parking meter and the red sign. There is no mention or indication of any yellow-plastic container in the image. 4. Finally, let's draw a conclusion: Since there is no mention of a yellow-plastic container in image 2, we cannot confirm its presence. Therefore, the correct answer is that there is no yellow-plastic container behind the grey parking meter. <answer>B</answer> (a) Baseline model (Qwen2.5-VL-3B) 1. Observing image 1, we can see two yellow-plastic containers standing side by side on a sidewalk. 2. Looking at image 2, we see a grey parking meter on a sidewalk with a yellow sign in the background. 3. Comparing image 2 to the scene in image 1, we can see that the yellow sign in image 2 matches the yellow container in image 1. 4. Therefore, based on the matching yellow sign and container, we can conclude that there is a yellow-plastic container behind the grey parking meter in image 2. 5. The answer is A. <answer>A</answer> (c) SpatialBlock-3B-reason Let's analyze the images step by step to determine if there is a yellow-plastic container behind the grey parking meter from the viewpoint presented in image 2. 1. Image 1: This image shows two yellow portable toilets (porta potties) on a sidewalk. There is no mention of a yellow-plastic container in this image. 2. Image 2: This image shows a close-up of a grey parking meter on a sidewalk. There is a red and white sign in the background, but no yellow-plastic container is visible in this image. 3. Image 3: This image shows a close-up of a yellow plastic container, which appears to be a porta potty. The container is positioned on a sidewalk, and there is a tree in the background. Based on the analysis: Image 1 does not show a yellow-plastic container. Image 2 does not show a yellow-plastic container. Image 3 shows a yellow-plastic container, which is a porta potty. From the viewpoint presented in image 2, there is no yellow-plastic container visible. Therefore, the answer is: <answer>B</answer> (b) SpatialBlock-3B-reason (without visual cue) Figure 9: Qualitative Comparison of Reasoning Traces. (a) Baseline only describes a single view. (b) Training without visual cues attempts multi-image reasoning but fails to identify the shared object. (c) Our model uses the āyellow containerā as an anchor to integrate views and reach the correct answer. Fig. 9 shows reasoning traces of the baseline model (Qwen2.5-VL-3B) and models trained on SpatialBlock-15k with and without visual cues for a question from the MindCube dataset. The task requires aggregating visual information from three images captured from different viewpoints. In addition, the model should infer the spatial relationship between two objects from the perspective presented in the second image (image 2). As shown in Fig. 9(a), the baseline model fails to exhibit multi-image understanding and merely describes image 2. The model trained without visual cues (Fig. 9(b)) attempts to reason over multiple images, but fails to identify the common object across views, leading to an incorrect conclusion. In contrast, SpatialBlock-3B-reason (Fig. 9(c)) identifies the āyellow containerā as an anchor object and connects the given images through this shared reference, ultimately producing the correct answer. These results demonstrate that training on our dataset enhances the modelās ability to per- form complex spatial reasoning. Furthermore, incorporating visual cues enables the model 19 to effectively leverage anchor objects to integrate information across multiple viewpoints. Additional qualitative results are illustrated in Fig. 10. To de te rmine the dire c tion of move me nt from the firs t vie w to the s e c ond vie w, le t's a na lyz e the give n informa tion: 1. **First View**: The image shows three trash bins arranged side by side. From left to right, they are: Blue recycling bin, Gray trash bin, Green trash bin with wheels 2. **Second View**: The image shows the green trash bin with wheels positioned against a wall. Given this information, we can deduce the following: The blue recycling bin is on the far left. The gray trash bin is in the middle. The green trash bin with wheels is on the far right. In the second view, the green trash bin with wheels is positioned against a wall, indicating that it has been moved from its original position next to the other bins to a new location. Now, let's consider the possible directions of movement: A. **Directly Right**: This would mean moving the green trash bin from its original position to the right, which aligns with the description of it being positioned against a wall. B. **Diagonally Forward and Right**: This would imply moving the green trash bin diagonally from its original position to theright, which does not match the description. C. **Directly Left**: This would mean moving the green trash bin from its original position to the left, which does not matchthe description. D. **Diagonally Forward and Left**: This would imply moving the green trash bin diagonally from its original position to the left, which does not match the description. Based on the analysis, the most logical direction of movement is directly right. Therefore, the correct answer is: <answer>A</answer> (a) Baseline model (Qwen2.5-VL-3B) 1. Observing view 1, we see three trash bins: a blue recycling bin, a gray trash bin, and a green trash bin with wheels. 2. In view 2, the green trash bin with wheels is the only one visible, and it appears to be in a different position. 3. The green trash bin with wheels in view 2 is positioned closer to the wall and appears to be in a different orientation compared to view 1. 4. Based on the change in the position and orientation of the green trash bin, it seems that the camera has moved diagonally forward and to the right. 5. Therefore, the direction from view 1 to view 2 is diagonally forward and right. <answer>B</answer> (c) SpatialBlock-3B-reason Let's analyze the two views step by step: 1. **View 1**: We see three trash bins: a blue recycling bin, a gray trash bin, and a green trash bin. There is also a blackboard with some stickers on it. 2. **View 2**: The green trash bin is now in a different position, suggesting that it has been moved. The other two bins (blue recycling bin and gray trash bin) are still in the same position, but the green trash bin is now closer to the camera. 3. **Direction Analysis**: From View 1, the green trash bin is to the right of the blue recycling bin and the gray trash bin.InView 2, the green trash bin is still to the right of the blue recycling bin and the gray trash bin, but it is now closer to the camera, indicating a movement in the direction of the camera. Based on this analysis, the direction of movement from View 1 to View 2 is diagonally forward and right. <answer>B</answer> (b) SpatialBlock-3B-reason (without visual cue) Q.Based on these two views showing the same scene: in which direction did I move from the first view to the second view? A. Directly right B. Diagonally forward and right C. Directly left D. Diagonally forward and left Image 1 Image 2 Figure 10: Additional Qualitative Result on Reasoning Traces. The figure shows additional qualitative results on reasoning traces of the baseline model, our model trained without visual cues, and our final model, SpatialBlock-3B-reason. 20