WorksComputer Science
GenDSR: Transferring Representations from Video Generation Models for Dynamic Spatial Reasoning
Abstract
Dynamic spatial reasoning is essential for connecting visual intelligence to thephysical world, yet remains challenging for vision-language models (VLMs), which must preserve scene state while tracking how spatial relations evolve over time. Meanwhile, video generation models (VGMs), trained to synthesize temporally coherent world dynamics, offer a rich source of spatiotemporal representations. We present GenDSR, a framework that transfers intermediate representations from VGM into VLM for dynamic spatial reasoning. Rather than treating a VGM feature as a monolithic prior, GenDSR constructs two complementary branches: a raw branch preserving holistic generative context and a delta branch emphasizing temporal variation. Task-conditioned tokenwise gates then fuse both signals and inject the result into the native visual representation. On DSR-Bench, GenDSR achieves 67.2%, surpassing the previous best result and establishing a new state of the art.
Introduction
The physical environments in which intelligent systems operate are constantly changing: objects move, viewpoints shift, and spatial relationships evolve over time. Reasoning in such environments requires more than recognizing static scene states—it requires tracking and understanding their changes across time. Dynamic spatial reasoning, which relates observed motion to changes in position, orientation, distance, and inter-object relations, is therefore fundamental to intelligence in the physical world. Tasks such as navigation, manipulation, and autonomous decision-making depend not only on understanding the current spatial configuration, but also on anticipating how it changes [25, 7].
Yet, despite substantial advances in vision-language models and video understanding, dynamic spatial reasoning remains far from reliable: recent benchmarks reveal persistent failures in integrating motion cues, disentangling camera and object motion, and maintaining temporally coherent spatial estimates [9, 24, 23].
A prominent line of recent work enhances VLMs with spatial priors from pretrained 3D foundation models [25, 20], such as VGGT [13] and [14]. While their latent features encode rich geometric structure, they are primarily learned for scene reconstruction rather than modeling temporal evolution, and rely on geometry-centric training data that are costly and difficult to collect at scale.
In contrast, video generation models [12, 18] provide a particularly suitable source of priors for dynamic spatial reasoning. Specifically, (1) Spatial consistency. To synthesize coherent videos under object and camera motion, these models must preserve scene structure and object identity across continuously changing viewpoints. Prior studies further suggest that their internal representations capture meaningful spatial cues, including camera pose and 3D structure, indicating spatial awareness beyond appearance-level semantics [8]. Shen et al., 2026). (2) Temporal dynamics. Video generation explicitly requires modeling how visual content evolves over time, including object motion, camera movement, and interactions among entities. This objective naturally encourages representations sensitive to temporal changes that are essential for reasoning about evolving spatial relationships [6]. (3) Scalable learning. Unlike geometry-centric models that depend on specialized multi-view data and geometric supervision, video generation models can learn directly from abundant real-world video collections. This provides a scalable source of both spatial and temporal knowledge without requiring costly 3D annotations or reconstruction labels. This naturally raises a question: can the representations learned by video generation models serve as dynamics-aware priors for spatial reasoning?
To this end, we propose GenDSR, a lightweight framework that transfers generative representations into VLMs for dynamic spatial reasoning. Our key insight is to expose two complementary views of the generative representation: its holistic spatiotemporal context and its explicit local variations. Accordingly, GenDSR constructs generative representations with two complementary branches. (1) Raw branch. We directly align the intermediate features of the video generation model with the VLM visual space, preserving rich spatial and spatiotemporal information about the scene. (2) Delta branch. We explicitly capture temporal variations by computing feature differences across neighboring observations, emphasizing changes induced by object motion, viewpoint shifts, and evolving spatial relationships. The two branches are jointly integrated through lightweight residual fusion, complementing the semantic representations of the original VLM with information about both spatial states and their temporal evolution.
On DSR-Bench [25], GenDSR improves the Qwen3-VL-8B SFT baseline from 62.0% to 67.2%, while the 4B variant improves from 58.8% to 62.7%. Raw and delta representations independently improve the accuracy, and their combination yields the strongest overall performance. These results validate video generation models as a promising source of dynamic spatial priors for vision-language reasoning.
Our contributions are summarized as follows:
We explore video generation models as a source of spatial and temporal priors for dynamic spatial reasoning, offering an alternative to geometry-centric representations.
We propose GenDSR, which constructs two complementary representations from generative features: a raw representation of spatial states and a delta representation emphasizing temporal variations, and efficiently integrates both into VLMs.
Experiments demonstrate consistent improvements in dynamic spatial reasoning and reveal the complementary roles of generative spatial context and temporal variations.
Related Work
Dynamic Spatial Reasoning
Spatial reasoning in vision-language models has progressed from understanding static layouts to reasoning about spatial configurations that evolve over time [16, 25]. Unlike static settings, dynamic spatial reasoning requires models to integrate evidence across frames, distinguish object motion from viewpoint changes, and maintain temporally coherent estimates of position, orientation, and inter-object relations. Recent studies show that even strong VLMs remain unreliable under these coupled spatial and temporal demands [24, 7].
Existing methods improve dynamic spatial reasoning through task-specific post-training, explicit geometry, or structured internal representations. SpaceR [10] combines large-scale video spatial data with spatially guided reinforcement learning. DSR Suite [25] formalizes dynamic spatial reasoning from monocular videos and introduces geometry-aware training data and a Geometry Selection Module that retrieves question-relevant information from pretrained 3D reconstruction priors. Building on geometry-centric representations, GeoSR [20] encourages VLMs to actively exploit geometry tokens through visual-token masking and adaptive geometric fusion. DynTrace [4] instead maintains geometry-informed object trajectories and structured dynamic traces at inference. Related work has also explored learning internal geometric or dynamic representations for reasoning. 3DThinker [2] aligns latent reasoning states with representations from a pretrained 3D foundation model, while
4DThinker Chen et al., [3] introduces dynamic latent imagery and reinforcement learning for spatiotemporal reasoning.
Video Generation Models for Spatial Reasoning
Recent evidence suggests that video generation models learn spatial structure beyond the appearance and motion statistics required for synthesis [8, 11]. Moving from analysis to transfer, VEGA-3D Wu et al., [15] extracts intermediate video-diffusion features as implicit 3D priors and adaptively fuses them with MLLM visual tokens for scene understanding. DyVA Zhang et al., [19] similarly repurposes a video diffusion model as a generative encoder, querying it with a single denoising step and injecting the resulting latents as auxiliary visual embeddings. GeoWorld-VLM Gu et al., [5] transfers geometry from a camera-conditioned world model through feature distillation. Another line of work turns generative models into explicit visual imagination. MindJourney Yang et al., [17] couples a VLM with a controllable world model at test time and iteratively synthesizes views along model-proposed camera trajectories. Zhang et al. [22] systematically studies this generate-then-reason paradigm. World2VLM Zhang et al., [21] instead uses trajectory-conditioned world-model generations as training-time supervision, distilling forward and inverse spatial reasoning into the VLM. Different from prior approaches that transfer generative representations as a unified auxiliary feature, GenDSR focuses on dynamic spatial reasoning and explicitly exposes two complementary views of the VGM representation: its raw spatiotemporal state and its local temporal variation. This decomposition allows the VLM to access temporal changes directly rather than relying on them to remain implicitly encoded in the raw generative feature.
Method
Overview
Given a video , a question , and an answer sequence , a video-language model first encodes the video into visual representations and predicts the answer autoregressively. We denote the frozen visual encoder by and the native visual merger by :
where denotes the number of visual tokens and is the hidden dimension of the language model. We associate these tokens with a spatiotemporal grid , such that . Our goal is to enrich the native visual representation with complementary knowledge from a frozen video generation model (VGM), while preserving the original visual pathway of the VLM.
Figure 1 provides an overview of GenDSR. We extract intermediate representations from the pretrained VGM and construct two complementary representations from them: a raw branch that preserves holistic generative context and a delta branch that explicitly emphasizes temporal variations. Both representations are aligned to the native VLM token grid and projected into the language-model hidden space. A lightweight, task-conditioned fusion module then determines their contributions at each visual-token position and injects the resulting representation as a residual immediately before the language model.
Figure 1. Overview of GenDSR. The frozen native visual encoder and video generation model process the same input video. We construct raw and delta representations from the VGM features and integrate them with the native visual representation through task-conditioned fusion before language-model reasoning. Snowflake and flame icons denote frozen and trainable modules, respectively.
The following sections introduce the raw–delta representation transfer in § 3.2, the adaptive fusion mechanism in § 3.3, and the training objective in § 3.4.
Raw–Delta VGM Representation Transfer
Generative feature extraction. Unlike a conventional visual encoder, a diffusion-based VGM operates through a latent denoising process. We first encode the observed video with the frozen VGM VAE:
For timestep , we draw and apply the native VGM scheduler’s noising operator :
where follows the VGM’s native noise schedule. We then perform one forward pass through the frozen denoising transformer under the empty-prompt condition . Let denote the intermediate activation at block , after patch embedding and propagation through the preceding transformer blocks. We extract
where the first three dimensions form the native VGM feature grid and is the VGM channel dimension. The empty-prompt condition prevents an external text prompt from contributing semantic information. This extraction requires only a single denoising-network forward at a fixed timestep, rather than an iterative denoising trajectory.
The VGM grid generally differs from the VLM grid . We use one shared, nonparametric alignment operator for both branches:
It interpolates the temporal axis to , applies adaptive average pooling to the target spatial grid , and flattens the aligned grid into tokens. The two branches use projectors with the same architecture but independent parameters. For , each projector is implemented as
where the first linear layer maps to dimensions and the second operates in the -dimensional language-model space.
Raw branch. The raw VGM representation retains the rich spatiotemporal context learned through video generation. We first align its native spatiotemporal grid with that of the VLM and then project the aligned features into the language-model hidden space:
where is a learnable feature projector. Rather than replacing the native representation , the resulting feature provides complementary spatial and spatiotemporal context acquired from generative pretraining.
Delta branch. While the raw representation contains rich contextual information, temporal changes may remain entangled with appearance and scene content. We therefore explicitly expose temporal variations in the VGM feature space by computing temporal differences between neighboring features:
where denotes the temporal resolution of the VGM representation. Prepending the zero feature preserves the original temporal length, so contains the same positions as . Importantly, temporal differencing is performed in the native VGM feature space before spatiotemporal alignment, so temporal differences are defined directly on the native VGM representation rather than on resampled features.
The resulting temporal differences are subsequently aligned and projected:
where is a learned projector. The delta representation emphasizes temporal variations associated with object motion, viewpoint changes, and evolving spatial relationships, complementing the holistic context retained by .
Task-Conditioned Tokenwise Fusion
The usefulness of raw and delta information can vary across questions and visual locations. For example, questions about static spatial configurations may rely primarily on scene context, whereas questions about motion or changing relationships may benefit more from temporal variations. A fixed global mixture therefore cannot adapt to the diverse requirements of dynamic spatial reasoning.
Let index the valid prompt-side text tokens, excluding visual tokens, target-response tokens, and padding, and let be the corresponding input embedding from the VLM language embedding layer. For our multiple-choice prompts, includes both the question and its answer-option text. We obtain the task representation by mean pooling:
The input embeddings are detached only when constructing , which prevents the gate from introducing an auxiliary gradient path through the prompt representation. We broadcast over the visual-token grid and condition the gate on the native visual tokens, both VGM branches, and the resulting task context. Specifically, we form
where denotes feature concatenation and are independent linear heads from to . The unconstrained coefficients allow each branch to be independently amplified, suppressed, or sign-adjusted. We initialize and to zero and their biases to and , respectively.
The fused generative residual is then computed as
where denotes tokenwise scaling with broadcasting over the hidden dimension. Only visual-token representations are modified; text-token embeddings remain unchanged. The residual is injected after the native visual merger and immediately before the language model, allowing the generative features to complement rather than replace the native VLM representation.
Training Objective
The augmented visual tokens together with the question condition the standard autoregressive supervised fine-tuning objective:
| Models | A.Dis | A.Dir | A.Ori | A.Spd | A.SpdC | A.DirP | R.Dis | R.Dir | R.Ori | R.Spd | R.SpdC | R.DirP | N-Temp | Avg. ↑ |
| Proprietary Models | ||||||||||||||
| GPT-5 | 21.1 | 41.5 | 48.7 | 34.5 | 33.3 | 34.7 | 17.2 | 44.3 | 41.9 | 21.2 | 25.0 | 30.9 | 26.7 | 30.8 |
| Gemini-2.5-Pro | 20.0 | 44.6 | 53.6 | 27.3 | 38.7 | 30.5 | 23.2 | 32.9 | 43.2 | 17.1 | 28.5 | 27.9 | 34.3 | 31.7 |
| Spatial Reasoning Models | ||||||||||||||
| VLM-3R | 28.2 | 27.6 | 31.7 | 42.8 | 38.7 | 33.0 | 34.4 | 23.8 | 30.8 | 22.2 | 26.4 | 29.1 | 35.0 | 31.4 |
| VG-LLM | 55.2 | 32.3 | 58.5 | 57.1 | 51.6 | 32.2 | 56.0 | 36.3 | 32.0 | 30.3 | 32.1 | 29.1 | 27.9 | 38.4 |
| DSR Suite-Model | 87.0 | 73.8 | 84.1 | 73.8 | 72.0 | 35.5 | 75.8 | 76.1 | 77.7 | 60.6 | 37.1 | 35.1 | 46.4 | 58.9 |
| GeoSR | 88.0 | 84.4 | 87.8 | 83.3 | 78.5 | 41.0 | 79.5 | 83.0 | 83.8 | 76.5 | 48.9 | 39.8 | 52.9 | 66.1 |
| 4DThinker | 84.3 | 79.7 | 82.3 | 84.0 | 67.0 | 47.8 | 67.0 | 82.8 | 79.5 | 62.9 | 44.2 | 41.0 | 48.5 | 62.0 |
| Ours (Based on Qwen3-VL-4B) | ||||||||||||||
| Qwen3-VL-4B | 34.1 | 31.7 | 29.5 | 37.5 | 14.3 | 30.4 | 22.5 | 29.5 | 35.4 | 20.8 | 19.6 | 20.6 | 35.7 | 27.7 |
| +SFT | 89.0 | 71.4 | 82.1 | 72.5 | 62.6 | 43.5 | 77.5 | 76.1 | 79.7 | 49.0 | 38.4 | 38.8 | 46.0 | 58.8 |
| +SFT+Raw | 89.0 | 73.0 | 85.9 | 68.8 | 65.9 | 46.1 | 77.5 | 78.4 | 82.3 | 59.4 | 41.3 | 33.3 | 48.1 | 60.2 |
| +SFT+Raw+Delta | 91.5 | 76.2 | 91.0 | 66.2 | 74.7 | 42.6 | 77.5 | 81.8 | 81.0 | 75.0 | 50.7 | 37.6 | 43.0 | 62.7 |
| Ours (Based on Qwen3-VL-8B) | ||||||||||||||
| Qwen3-VL-8B | 23.5 | 24.6 | 42.6 | 29.7 | 27.9 | 33.8 | 18.1 | 28.4 | 34.5 | 24.2 | 22.1 | 27.9 | 33.5 | 28.7 |
| +SFT | 89.0 | 82.5 | 88.5 | 78.8 | 70.3 | 47.8 | 74.8 | 83.0 | 86.1 | 63.5 | 42.0 | 36.4 | 43.4 | 62.0 |
| +SFT+Raw | 89.0 | 87.3 | 89.7 | 77.5 | 80.2 | 43.5 | 76.6 | 88.6 | 86.1 | 83.3 | 58.7 | 38.2 | 41.3 | 65.8 |
| +SFT+Raw+Delta | 89.0 | 82.5 | 92.3 | 77.5 | 80.2 | 44.3 | 76.6 | 90.9 | 83.5 | 84.3 | 53.6 | 38.2 | 52.0 | 67.2 |
Table 1. Dynamic spatial reasoning on DSR-Bench. Accuracy (%) is reported for 13 subtasks and overall accuracy.
where denotes the length of the target answer and contains all trainable parameters. Both the VGM, including its VAE and denoising transformer, and the native VLM visual encoder remain frozen. We optimize the language model, the native visual merger, the two branch projectors, and the fusion module.
Experiments
Experimental setup
Benchmark and metrics. We evaluate on DSR-Bench [25], a multiple-choice benchmark for fine-grained dynamic spatial reasoning from monocular videos. It contains 13 subtasks: six spatial attributes—distance, direction, orientation, speed, speed comparison, and direction prediction—are evaluated under absolute and relative viewpoints, together with a non-template subset. We report accuracy for each subtask and overall sample-level accuracy.
Implementation details. Our default model uses Qwen3-VL-8B-Instruct [1] as the VLM and frozen Wan2.1-T2V-1.3B [12] as the VGM. We extract VGM activations offline from zero-indexed block 20 under an empty-text condition and diffusion timestep of 300. We train on the DSR-train split for one epoch with a global batch size of 32 and a learning rate of . The VGM and native VLM visual encoder remain frozen. We use AdamW with a cosine learning-rate schedule, a warmup ratio of 0.03, and weight decay of 0.01. All experiments are conducted on 8×A100 GPUs.
Results and Analysis
Table 1 evaluates whether VGM representations improve over strong public and proprietary systems. GenDSR with Qwen3-VL-8B reaches 67.2%, surpassing the previous best GeoSR [20], which uses the 3D foundation model [14], establishing a new state of the art on DSR-Bench. The gain is not explained by supervised fine-tuning alone: Qwen3-VL-8B SFT obtains 62.0%, raw transfer raises accuracy to 65.8%, and the full raw–delta method reaches 67.2%. The 4B backbone shows the same ordering, with raw–delta fusion reaching 62.7% compared with 58.8% for SFT. These cross-scale results show that the transferred representation improves both backbones under the protocol.
Ablation studies
Raw and delta branches. Table 2 evaluates the two transferred signals. Raw and delta features independently improve over SFT, while their combination performs best, supporting their complementarity.
| Variant | Raw | Delta | Acc |
| Base | – | – | 28.70 |
| SFT | – | – | 62.00 |
| Raw only | ✓ | – | 65.80 |
| Delta only | – | ✓ | 64.32 |
| GenDSR | ✓ | ✓ | 67.20 |
Table 2. Raw–delta components.
VGM feature block. Table 3 varies the extraction layer while keeping the raw–delta pathway, diffusion timestep, and training protocol fixed. Block 20 performs best, while both shallower and deeper features are weaker. This indicates that transferable evidence is concentrated in an intermediate-to-late representation rather than increasing monotonically with network depth.
| Block | Acc |
| 5 | 65.46 |
| 10 | 65.39 |
| 15 | 66.64 |
| 20 | 67.20 |
| 29 | 66.43 |
Table 3. VGM block.
Diffusion timestep. Table 4 examines how the diffusion timestep shapes the transferred representation. Performance is lower for the clean and lightly noised states, peaks at an intermediate timestep, and then gradually declines as the corruption becomes stronger. This non-monotonic trend suggests that an intermediate noising state may better balance the preservation of input evidence with the VGM’s learned denoising prior.
| Timestep | Acc |
| No noise | 65.60 |
| 65.04 | |
| 67.20 | |
| 66.92 | |
| 66.43 |
Table 4. VGM timestep.
Conclusion
We introduce GenDSR, which extracts intermediate representations from a frozen video generation model and transfers them into a VLM through complementary raw and delta branches. Experiments on DSR-Bench show that generative representations consistently improve VLM performance. These results suggest that video generation models can serve not only as synthesis models, but also as a valuable source of representations for reasoning about the evolving physical world.
References
- [1]Shuai Bai, Yuxuan Cai, Ruizhe Chen, Kegin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Lu, Shixuan Lu, Dunjie Lu, Ruilin Luo, Chenxu Lv, Rui Men, Lingchen Meng, Xuancheng Ren, Xingzhang Ren, Sibo Song, Yuchong Sun, Jun Tang, Jianhong Tu, Jianqiang Wan, Peng Wang, Pengfei Wang, Qiuquye Wang, Yuxuan Wang, Tianbao Xie, Yiheng Xu, Haiyang Xu, Jin Xu, Zhibo Yang, Mingkun Yang, Jianxin Yang, An Yang, Bowen Yu, Fei Zhang, Hang Zhang, Xi Zhang, Bo Zheng, Humen Zhong, Jingren Zhou, Fan Zhou, Jing Zhou, Yuanzhi Zhu, and Ke Zhu. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025.arxiv.org/abs/2511.21631
- [2]Zhangquan Chen, Manyuan Zhang, Xinlei Yu, Xufang Luo, Mingze Sun, Zihao Pan, Yan Feng, Peng Pei, Xunliang Cai, and Ruqi Huang. Think with 3d: Geometric imagination grounded spatial reasoning from limited views. arXiv preprint arXiv:2510.18632, 2025.arxiv.org/abs/2510.18632
- [3]Zhangquan Chen, Manyuan Zhang, Xinlei Yu, Xiang An, Bo Li, Xin Xie, ZiDong Wang, Mingze Sun, Shuang Chen, Hongyu Li, et al. 4dthinker: Thinking with 4d imagery for dynamic spatial understanding. arXiv preprint arXiv:2605.05997, 2026.arxiv.org/abs/2605.05997
- [4]Rongxin Gao, Yuzhi Huang, Dongxuan Liu, Chu Li, Zhenye Wang, Jie Wu, Shuzhao Xie, Jingyan Jiang, Xinghao Ding, Xiaotong Tu, and Yue Huang. Dyntrace: Tracking dynamic object evidence for 4d spatio-temporal reasoning in mlms, 2026.arxiv.org/abs/2607.12503
- [5]Renjie Gu, Kaichen Zhou, Yan Luo, and Mengyu Wang. Geoworld-vlm: Geometry from world models for vision-language models, 2026.arxiv.org/abs/2605.16713
- [6]Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022.arxiv.org/abs/2204.03458
- [7]Yuzhi Huang, Kairun Wen, Rongxin Gao, Dongxuan Liu, Yibin Lou, Jie Wu, Jing Xu, Jian Zhang, Zheng Yang, Yunlong Lin, et al. Thinking in dynamics: How multimodal large language models perceive, track, and reason dynamics in physical 4d world. arXiv preprint arXiv:2603.12746, 2026a.arxiv.org/abs/2603.12746
- [8]Zixuan Huang, Xiang Li, Zhaoyang Lv, and James M Rehg. How much 3d video foundation models encode? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 384–394, 2026b.arxiv.org/abs/2512.19949
- [9]Chih-Ting Liao, Xi Xiao, Chunlei Meng, Zhangquan Chen, Yitong Qiao, Weilin Zhou, Tianyang Wang, Xu Zheng, and Xin Cao. Spamem: Benchmarking dynamic spatial reasoning via perception-memory integration in embodied environments, 2026.arxiv.org/abs/2604.22409
- [10]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.arxiv.org/abs/2504.01805
- [11]Haozhan Shen, Tiancheng Zhao, Kangjia Zhao, and Jianwei Yin. Which pretraining paradigm better serves spatial intelligence? an empirical comparison of vision-language and video generation models. arXiv preprint arXiv:2605.28132, 2026.
- [12]Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025.
- [13]Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5294–5306. IEEE, 2025.arxiv.org/abs/2503.11651
- [14]Yifan Wang, Jianjun Zhou, Haoyi Zhu, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Jiangmiao Pang, Chunhua Shen, and Tong He. π³: Permutation-equivariant visual geometry learning. In International Conference on Learning Representations, pages 10481–10497, 2026.
- [15]Xianjin Wu, Dingkang Liang, Tianrui Feng, Kui Xia, Yumeng Zhang, Xiaofan Li, Xiao Tan, and Xiang Bai. Generation models know space: Unleashing implicit 3d priors for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), 2026.arxiv.org/abs/2603.19235
- [16]Jihan Yang, Shusheng Yang, Anjali Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in Space: How Multimodal Large Language Models See, Remember and Recall Spaces. In CVPR, 2025a.
- [17]Yuncong Yang, Jiageng Liu, Zheyuan Zhang, Siyuan Zhou, Reuben Tan, Jianwei Yang, Yilun Du, and Chuang Gan. Mindjourney: Test-time scaling with world models for spatial reasoning, 2025b.
- [18]Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer. In International Conference on Learning Representations, pages 83048–83077, 2025c.
- [19]Kevin Zhang, Kuangzhi Ge, Xiaowei Chi, Renrui Zhang, Shaojun Shi, Zhen Dong, Sirui Han, and Shanghang Zhang. Can world models benefit vlms for world dynamics?, 2025a.
- [20]Shihua Zhang, Qiuhong Shen, Shizun Wang, Tianbo Pan, and Xinchao Wang. Make geometry matter for spatial reasoning. In European Conference on Computer Vision, pages 231–249. Springer, 2026a.arxiv.org/abs/2603.26639
- [21]Wanyue Zhang, Wenxiang Wu, Wang Xu, Jiaxin Luo, Helu Zhi, Yibin Huang, Shuo Ren, Zitao Liu, and Jiajun Zhang. World2vlm: Distilling world model imagination into vlms for dynamic spatial reasoning, 2026b.
- [22]Yanbing Zhang, Bo Wang, Jianhui Liu, Nan Jiang, Jiaxiu Jiang, Haoze Sun, Yijun Yang, Shenghe Zheng, Lin Song, Haoyang Huang, Nan Huan, and Wenbo Li. Thinking with novel views: A systematic analysis of generative-augmented spatial intelligence, 2026c.arxiv.org/abs/2605.10588
- [23]Ziang Zhang, Zehan Wang, Guanghao Zhang, Weilong Dai, Yan Xia, Ziang Yan, Minjie Hong, and Zhou Zhao. Dsi-bench: A benchmark for dynamic spatial intelligence. arXiv preprint arXiv:2510.18873, 2025b.
- [24]Shijie Zhou, Alexander Vilesov, Xuehai He, Ziyu Wan, Shuwang Zhang, Aditya Nagachandra, Di Chang, Dongdong Chen, Xin Eric Wang, and Achuta Kadambi. Vlm4d: Towards spatiotemporal awareness in vision language models. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8600–8612. IEEE, 2025.
- [25]Shengchao Zhou, Yuxin Chen, Yuying Ge, Wei Huang, Jiehong Lin, Ying Shan, and Xiaojun Qi. Learning to reason in 4d: Dynamic spatial understanding for vision language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9637–9646, 2026.
