SpatialVTS: Vision-Language Spatial Reasoning

• SpatialVTS is a vision-language method that integrates explicit visual grounding with textual reasoning to improve spatial understanding in both 2D and 3D images.
• It employs a two-phase pipeline—Spatial Visual Thinking for region token prediction and Spatial Textual Thinking for generating detailed reasoning traces and answers.
• The approach achieves competitive performance on both qualitative and quantitative benchmarks while eliminating the need for additional inputs like depth or masks.

SpatialVTS is a vision-language method for spatial reasoning whose name expands to “Spatial reasoning through Visual and Textual thinking Simultaneously.” It is designed to improve spatial reasoning in 2D and 3D space by coupling explicit visual grounding of task-relevant regions with long-form textual reasoning, while operating with standard RGB images and text only, without additional inputs such as masks or depth. The method is organized as a two-phase pipeline—Spatial Visual Thinking and Spatial Textual Thinking—and is trained on a reconstructed dataset derived from SpatialRGPT and VPT, with manual correction of noisy labels and addition of reasoning traces (Liang et al., 28 Jul 2025).

1. Definition and problem setting

SpatialVTS addresses image-based spatial reasoning tasks that require more than object recognition. The target problem class includes qualitative spatial reasoning, such as left/right, up/down, and front/behind, and quantitative spatial reasoning, such as distance, height, width, and scale. The paper frames these capabilities as fundamental to visual question answering (VQA) and robotics, but the reported experiments are conducted on image-question-answer benchmarks rather than on interactive or embodied control tasks (Liang et al., 28 Jul 2025).

The method is motivated by a specific diagnosis of current VLM failure modes. Spatial questions are often not answered by directly visible labels alone; they require identifying the correct visual references, sometimes including auxiliary objects not explicitly named in the question, and then performing multi-step inference over those references. The paper argues that short-answer supervision, such as training only on outputs like “left” or “yes,” can encourage superficial answer patterns or hallucinations rather than genuine spatial reasoning. SpatialVTS therefore treats spatial reasoning as a compound process in which visual evidence selection and textual inference must be improved together.

A central distinction in the paper is between objects explicitly mentioned in the question and potential objects related to the reasoning. This matters because some spatial judgments depend on contextual cues or reference structures. The method is explicitly designed to retrieve not only evident targets but also latent supports for inference. This differentiates SpatialVTS from approaches that ground only named entities.

2. Method architecture

SpatialVTS is organized into two phases: Spatial Visual Thinking and Spatial Textual Thinking. In the first phase, the model predicts location-related specific tokens for spatially relevant regions. In the second phase, it uses the question together with cropped visual cues from those regions to generate rationales and a final answer (Liang et al., 28 Jul 2025).

The visual-thinking stage uses a discrete region representation rather than direct box regression. An image of size $h \times w$ is evenly divided into a $k \times k$ grid, with each cell having size

$\frac{h}{k} \times \frac{w}{k}.$

In implementation, the paper sets

$k = 8.$

A region is represented by the cells containing its top-left and bottom-right pixels, with token sequences such as

$<x_0><y_3><x_3><y_6>.$

This discretization is intended to make location prediction easier to train while preserving coarse spatial structure.

The paper emphasizes that the first stage is not restricted to question-mentioned objects. SpatialVTS is trained to produce tokens for essential targets automatically, including potential targets related to the reasoning. Those regions are then cropped from the original image and reused as visual cues. The candidate objects are cropped as new images and input into the same vision encoder, rather than being processed by a separate visual feature extractor.

The textual-thinking stage then performs long-form reasoning over the original question and the cropped cues. The paper describes this as reasoning based on visual cues and dialogues; in practice, the provided description indicates question/context text together with generated rationale traces. The rationale is intended to establish connections among the selected targets and to avoid the pattern-collapse problem associated with training only on short answers. This stage is therefore still multimodal, but the supervision is explicitly textual in the form of reasoning traces.

Two contrasts with earlier systems are central. Relative to VPT, SpatialVTS includes potential reasoning clues, not only objects directly named in the problem. Relative to SpatialRGPT, it avoids dependence on depth and mask inputs, keeping the formulation within a more general RGB-plus-text VLM regime.

3. Dataset reconstruction, supervision, and training

A substantial part of SpatialVTS is its reconstruction of training data. The training set is built from 100k samples from the original SpatialRGPT dataset and 340k samples from VPT, including 140k spatial relation questions and 200k general questions (Liang et al., 28 Jul 2025).

The paper motivates reconstruction on three grounds. First, the original input format of SpatialRGPT depended on specialized placeholders such as <mask> and <depth>, which the authors regarded as too specialized for generalized image-text spatial reasoning. Those placeholders are replaced by plain-text labels such as Region [0] and Region [1], and region boxes are drawn on the original images. Second, the authors report that the automatically constructed source data contained a considerable proportion of incorrect answers, many unreasonable question settings, and questions uncorrelated with the images. They therefore manually filter unreasonable examples and manually correct incorrect answers. Third, the original data lacked process-oriented thinking, so the reconstructed corpus adds potential target boxes, evidence/rationales explaining why each target matters, and long-form reasoning traces.

Reasoning traces are generated with what the paper calls “seeking the cause by grasping the result.” During rationale construction, the system provides a public VLM with the question, the visual cues, and the known correct answer, and asks it to generate a reasoning path that leads to that answer. The exact public VLM used for this rationale generation is not specified in the provided description.

The backbone is Qwen2-VL, and the reported setup is effectively a 7B-class VLM. Training uses 8 NVIDIA H800 GPUs and takes 7 to 10 hours. The paper presents two principal variants: SpatialVTS(vision), which omits textual chain-of-thought-style reasoning, and SpatialVTS(vision/text), which includes both stages.

4. Evaluation protocol and empirical results

SpatialVTS is evaluated on qualitative benchmarks—WhatsUP, VSR, BLINK-Spatial, and SRGPT-QUAL—and on quantitative benchmarks—Q-Spatial++ and SRGPT-QUAN (Liang et al., 28 Jul 2025). For qualitative evaluation, the paper uses DeepSeek-V3-0324 as an evaluator model; answers receive scores in $[0,1]$ and are accepted if the score is greater than 0.5. For quantitative evaluation, the paper defines

$$\delta=\max\left(\frac{\hat{d}}{d^*},\frac{d^*}{\hat{d}}\right),$$

and accepts predictions satisfying

$\delta \leq 2.$

One ablation also uses $\delta_{\leq 1.25}$.

On the qualitative benchmarks, SpatialVTS(vision/text) reports the best overall average score, 82.03. The same variant achieves 78.21 on VSR, 78.32 on SpatialRGPT-QUAL, 95.38 on WhatsUp, and 76.22 on BLINK-spatial. The vision-only variant, SpatialVTS(vision), reports an average of 75.96 and attains 85.31 on BLINK-spatial, indicating that the visual-thinking stage alone already improves grounding. The strongest non-SpatialVTS average baseline in the reported table is VPT at 74.47, while Qwen2.5-VL-7B reports 67.26 average (Liang et al., 28 Jul 2025).

The quantitative results are described more selectively in the provided record. The paper states that SpatialVTS outperforms all VLMs trained only with text-image pairs and is nearly on par with SpatialRGPT, despite SpatialRGPT using additional depth and segmentation masks. This is one of the central claims of the work, because SpatialVTS aims to obtain strong spatial reasoning using only RGB and text.

Ablation studies isolate the role of textual thinking. On SpatialRGPT-QUAN with $\delta_{\leq 1.25}$, removing textual thinking yields 34.78, while adding it yields 37.85. On Q-Spatial++ with $k \times k$0, the corresponding numbers are 48.10 without textual thinking and 54.45 with it. The paper summarizes these as improvements of roughly 3% and 7%, respectively. Another ablation compares a smaller reconstructed dataset of 250k samples, including 120k CoT samples, against the fuller setting. On VSR, the smaller setting reports 72.89 versus 78.21 for the fuller one; on Q-Spatial++, it reports 51.02 versus 54.45. The authors interpret this as evidence for both the value of the reconstructed dataset and the utility of additional CoT supervision.

5. Position within the spatial reasoning literature

SpatialVTS belongs to a broader research effort that distinguishes perception, reasoning, and action in spatial intelligence, but it occupies a specific niche within that landscape. It is a method for image-based spatial question answering, not an interactive agent benchmark, not a simulator-grounded repair loop, and not a 3D scene-memory model.

Its nearest conceptual antecedents in the provided corpus are VPT and SpatialRGPT, because these works also target spatial reasoning from visual inputs. SpatialVTS differs from VPT by explicitly modeling potential reasoning clues beyond mentioned objects and by reusing cropped candidate regions as evidence in a second-stage multimodal reasoning step. It differs from SpatialRGPT by avoiding depth and mask inputs and by converting the task into a generalized image-text format (Liang et al., 28 Jul 2025).

The benchmark literature clarifies the type of problem SpatialVTS does and does not solve. VSP decomposes visual spatial planning into perception and reasoning subtasks and shows that even strong VLMs fail on end-to-end planning, with large gains when images are replaced by structured text (Wu et al., 2024). iVISPAR moves further toward agentic evaluation by requiring iterative action selection in a formally analyzable sliding-geom puzzle environment with synchronized 2D, 3D, and text interfaces (Mayer et al., 5 Feb 2025). SpatialAct extends this action-conditioned perspective to simulator-grounded 3D scene refinement and reports a clear reasoning-to-action gap in multi-turn interaction (Liu et al., 29 May 2026). Relative to these benchmarks, SpatialVTS addresses the earlier stage of the pipeline: extracting the right visual evidence and performing the right textual inference on static image-based tasks.

Other contemporary systems pursue more explicit spatial representations. Spa3R learns a unified, view-invariant spatial representation from unposed multi-view images and injects a compact scene latent into a VLM through a cross-attention adapter (Jiang et al., 24 Feb 2026). GST-VLA replaces plain 2D patch tokens with structured Gaussian spatial tokens and supervises explicit intermediate 3D thoughts for action generation (Sarowar et al., 10 Mar 2026). These works indicate a different design philosophy from SpatialVTS: rather than augmenting a VLM with better evidence selection and rationale supervision, they build explicit spatial latent structures into the model’s input or control stack. This suggests that SpatialVTS is best understood as an RGB-only reasoning enhancement method, not as a full 3D world-model or action-policy architecture.

6. Limitations, interpretation, and significance

The paper does not include a dedicated limitations section, but several limitations are explicit or implicit in the reported design. First, quantitative spatial estimation remains hard from a single RGB image. The paper states that depth-assisted methods like SpatialRGPT still have an advantage for precise distance judgment. Second, benchmark ambiguity and answer underspecification can affect evaluation. Third, the experiments concern image-question-answer benchmarks, not embodied interaction, continuous control, or scene editing. SpatialVTS therefore should not be conflated with methods for simulator-grounded planning, video generation with persistent 3D memory, or SLAM-based spatial control.

A common misconception is to treat SpatialVTS as a generic “spatial VLM” in the broadest sense. The broader literature uses related ideas for very different purposes: Spatia uses an updatable 3D scene point cloud as persistent memory for long-horizon video generation (Zhao et al., 17 Dec 2025), while VTGaussian-SLAM uses view-tied 3D Gaussians for large-scale RGB-D SLAM (Hu et al., 3 Jun 2025). SpatialVTS is narrower. It is neither a video-generation framework nor a SLAM system; its contribution lies in coupling spatially informed region discovery with rationale-based textual reasoning under a standard RGB-plus-text interface.

Within that narrower scope, its significance is twofold. Methodologically, it argues that spatial reasoning in VLMs can be improved without introducing privileged modalities, provided that the model is taught to identify the right visual evidence and to reason over it explicitly. Empirically, it shows that this combination materially improves both qualitative and quantitative spatial understanding benchmarks, with the strongest reported qualitative average reaching 82.03 and quantitative performance approaching a depth- and mask-assisted model while using only RGB and text (Liang et al., 28 Jul 2025). This places SpatialVTS as a representative approach in the line of research that treats spatial reasoning not as a single inference step, but as the coordinated result of visual grounding, evidence selection, and textual deliberation.