The dominant spatial reasoning benchmarks for vision-language models focus on perception: extracting relations that are already visible in a single input image. Real-world spatial intelligence requires more than passive observation; it demands predictive reasoning—constructing a scene from limited observations, anticipating how an intervention changes it, and drawing conclusions about outcomes that are not directly visible. Benchmarks that stop at perception leave a gap in evaluating whether a model can perform the transformative and inferential steps that underpin autonomous navigation, manipulation planning, and interactive simulation. SpaceCast-Bench addresses this gap by introducing the first diagnostic evaluation of predictive spatial reasoning in vision-language models.
SpaceCast-Bench is built around an observe-transform-infer framework. The benchmark contains 3,862 questions derived from 182 real-world scenes. These questions span 16 task types organized into three difficulty levels. Static perception tasks require reading visible relations from the input. Local prediction tasks ask the model to anticipate how a localized change, such as moving an object, alters the immediate scene. Global prediction tasks require relational inference over unobserved outcomes, such as predicting the final state of a scene after a sequence of moves. Each level progressively demands more sophisticated scene understanding, spatial state updating, and reasoning about outcomes that were never directly observed.
The evaluation encompassed 21 vision-language models, including both general-purpose and spatially specialized variants. The strongest model achieved 58.0% accuracy across the benchmark, while human participants scored 87.2%. Spatially specialized models performed near random chance, indicating that the predictive reasoning demand is not adequately met by architectural biases alone. This result establishes a significant performance gap and underscores the difficulty of the predictive reasoning task.
Controlled analyses within the paper identify two factors that substantially affect model performance. First, bridge views—additional observations from different vantage points—are critical for integrating distributed information. Models that can incorporate multiple perspectives substantially outperform those relying on a single view. Second, explicit 3D evidence, such as depth maps or reconstructed geometry, benefits models more reliably than generated outcome images or videos. The authors attribute this to the fidelity and consistency of explicit 3D representations compared to synthesized visual approximations.
To demonstrate the benchmark's utility for model improvement, the authors fine-tuned Qwen3-VL-4B on programmatically generated SpaceCast-Bench data. After fine-tuning, the model's accuracy rose from 34.0% to 65.7%. Moreover, the fine-tuned model showed macro-average gains across six out-of-domain benchmarks, suggesting that the reasoning skills acquired on SpaceCast-Bench transfer to related spatial tasks. This result indicates that the benchmark not only evaluates predictive spatial reasoning but also provides a viable path for improving model capability in this domain.
Why Predictive Spatial Reasoning Matters for Developers
Vision-language models are increasingly deployed in systems that must operate in physical or simulated environments. Autonomous vehicles must reason about how moving an obstacle changes the drivable space. robotic manipulators must predict the result of grasping and relocating an object. Interactive agents must anticipate the consequences of user actions. Benchmarks that only test perception cannot evaluate whether a model possesses the forward-looking spatial intelligence required for these tasks. SpaceCast-Bench fills this gap by providing a structured, reproducible way to measure and improve a model's ability to reason about unobserved outcomes.
Benchmark Design: Observe-Transform-Infer Framework
The core design of SpaceCast-Bench is the observe-transform-infer cycle. In the observe phase, the model is presented with an initial scene, typically from a single camera view or a limited set of observations. The transform phase introduces an intervention—such as moving an object, rotating a component, or changing a spatial configuration. The infer phase requires the model to predict the resulting scene state, answering questions that are agnostic to the initially observed details but depend on the introduced change. This framework ensures that successful performance requires all three competencies: perception, transformation, and inference.
The 16 task types are derived from common spatial operations such as object relocation, occlusion reasoning, perspective transformation, and relational rearrangement. Each question is paired with a ground-truth answer derived from physically simulated scene evolution, providing a deterministic target for model predictions. The benchmark's programmatic generation pipeline ensures that questions can be instantiated at scale while preserving control over difficulty factors such as the number of occluding objects, the complexity of the transformation, and the depth of relational inference required.
Evaluation Results and Model Analysis
The 21 models evaluated span a range of architectures, including cross-attention-based vision-language transformers, hybrid CNN-attention systems, and encoder-decoder models. The top-performing model reached 58.0%, a substantial margin below the human baseline of 87.2%. Many models showed marked drops in accuracy when transitioning from static perception to local and global prediction, confirming that the predictive components of the benchmark are the primary source of difficulty.
The analysis of bridge views revealed that models incorporating multiple observation angles before the transform phase consistently outperformed single-view baselines. The gain was most pronounced in global prediction tasks, where integrating distributed observations is necessary to track object trajectories and relational shifts across the scene. Models that relied on generated outcome images or videos as evidence showed higher variance and lower average accuracy compared to models that received explicit 3D geometry, such as depth maps or voxel grids. The authors note that explicit 3D evidence provides a more compact and unambiguous representation of spatial relationships, which reduces the burden on the model's generative priors.
Fine-Tuning and Transfer Effects
The fine-tuning experiment with Qwen3-VL-4B is the most concrete demonstration of the benchmark's practical value. Starting from a baseline of 34.0% on SpaceCast-Bench, the model was trained on the programmatic question-answer pairs for a fixed number of steps. After fine-tuning, accuracy climbed to 65.7%, crossing the majority threshold. More notably, the authors report macro-average improvements across six out-of-domain benchmarks that test related spatial and reasoning capabilities. These gains suggest that the model learned reusable spatial transformation and inference heuristics rather than overfitting to the specific question formats in SpaceCast-Bench.
Limitations and Remaining Challenges
The authors acknowledge several limitations. The benchmark relies on physically simulated scenes, which may not capture all the variability of real-world environments such as lighting changes, texture effects, or dynamic agent interactions. The 182 source scenes were curated to cover a broad range of spatial configurations, but additional domains—such as crowded urban scenes or deformable object manipulations—may require extensions. The fine-tuning results, while promising, are based on a single model architecture; whether other models exhibit similar transfer effects remains an open question. Finally, the bridge-view and 3D evidence analyses are based on controlled experiments; real-world deployments may present additional constraints such as bandwidth limitations for transmitting multiple views or computational costs for 3D reconstruction.
Practical Takeaways for Model Development
For developers working on vision-language models, SpaceCast-Bench offers a concrete diagnostic and improvement pathway. The benchmark's structure—observation, transformation, inference—maps naturally to pipeline design decisions. Incorporating multiple camera perspectives or bridge views during both training and inference can measurably boost predictive accuracy. Prioritizing explicit 3D representations, such as depth estimation or meshes, over generated visual approximations can yield more reliable spatial reasoning. The fine-tuning protocol demonstrates that targeted training on predictive spatial tasks can transfer to unrelated benchmarks, suggesting that a dedicated phase of spatial reasoning training may be a worthwhile addition to model development curricula.
SpaceCast-Bench is released with question sets, simulation pipelines, and evaluation scripts to enable reproducible research and model comparison. The authors encourage the community to build upon the benchmark by extending the scene corpus, adding new task types, and investigating architectural modifications specifically tailored to predictive spatial reasoning.