Ascending the Infinite Ladder: Benchmarking Spatial Deformation Reasoning in Vision-Language Models

Under review. (given the final state, determine the operations). We adopt a ladder competition format, using the number of deformation steps as the level classification criterion, with the goal of exploring the boundaries of the model's deformation reasoning capabilities. Interestingly, the benchmarking results reveal that almost no model demonstrates plausible spatial deformation reasoning abilities. Furthermore, even after applying targeted training and mainstream reasoning enhancement methods, the models are still unable to perform well on 3D spatial deformation reasoning. Spatial Deformation Reasoning In this study, spatial deformation reasoning refers to the model's ability to understand, predict, and execute complex deformations of an object's shape. We focus on implementing this reasoning in VLMs, especially when there is no prior knowledge, and the model must learn and perform deformations through observation or manipulation. Unlike visual spatial intelligence in the VSI-Bench [56] , which emphasizes localization, relational understanding, and spatial perception, our spatial deformation centers on dynamic shape transformations, particularly in multi-step processes that change an object's state. We categorize the core capabilities of spatial deformation as follows: Spatial Recognition. The ability to comprehend the initial shape of an object and accurately identify the areas that require deformation. Abstraction of Operational Law Principles. The ability to grasp the principles behind deformation operations and abstract them into deformation laws, crucial for accurate reasoning. Stable Reasoning Execution. The ability to gradually execute the inferred reasoning rules, maintain stability throughout the process, and derive the final state.