Paper 2606.03988
Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models
- Published
- Jun 2026
- Research lab
- Independent
- Citations
- 0
- GitHub
- 102 stars
01 In brief
Summary
The paper introduces Imaginative Perception Tokens (IPTs), intermediate visual representations that externalize what a VLM would perceive under an alternative spatial configuration, to improve spatial reasoning.
Three tasks requiring imaginative perception are formulated: Perspective Taking (PET), Path Tracing (PT), and Multiview Counting (MVC), with datasets of ~20K examples each from simulated and real-world settings, paired with ground-truth intermediate imaginations and final answers.
Using the unified VLM BAGEL as backbone, IPT supervision improves spatial reasoning across settings, often outperforming textual chain-of-thought training, even without generating images at inference.
On MVC, IPT improves accuracy by 3.4% and achieves competitive performance on PT.
Mixed training with IPT and label-only data further improves performance.
Textual chain-of-thought can be detrimental, highlighting a modality mismatch.
IPT provides a principled supervision signal for reasoning over unobserved structure, yielding stronger spatial generalization and more interpretable intermediates aligned with task geometry.
Code will be released.
02 From the paper
Abstract
Vision language models (VLMs) excel at many tasks but still struggle with spatial reasoning when critical information is not directly observable. Many such problems require imaginative perception: inferring what would be seen from an unseen viewpoint, tracing paths through occluded spaces, or integrating partial observations into a coherent spatial representation. We introduce Imaginative Perception Tokens (IPT), intermediate perceptual representations that externalize what a VLM would perceive under alternative spatial configurations while remaining consistent with the observed input. To study this capability, we formulate three tasks, Perspective Taking (PET), Path Tracing (PT), and Multiview Counting (MVC), and construct datasets of approximately 20K examples with ground truth imaginations, answers, and evaluation benchmarks. Using the unified VLM BAGEL as the backbone, IPT supervision consistently improves spatial reasoning and often outperforms textual chain of thought training, even without generating images at inference time. On MVC, IPT improves accuracy by 3.4% and achieves competitive performance with strong closed-source models on PT. We further find that combining IPT and label-only supervision yields additional gains, whereas textual chain of thought can substantially degrade performance, suggesting a modality mismatch when spatial computation is forced through language. Overall, IPT provides a principled supervision signal for reasoning about unobserved spatial structure, improving generalization while producing interpretable intermediate representations.