Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models

2026-06-02

AI Summary

This work addresses the limitations of multimodal language models in spatial reasoning tasks—such as perspective transformation, occluded path tracing, and multi-view integration—where performance is hindered by the invisibility of critical information. To mitigate modality misalignment, the authors propose Imaginative Perceptual Tokens (IPTs), interpretable intermediate representations consistent with input observations that model spatial structures from unobserved viewpoints. Integrated within the unified vision-language model architecture BAGEL, IPTs introduce supervisory signals without requiring image generation, significantly enhancing reasoning capabilities. The approach improves accuracy by 3.4% on the Multiview Counting task, achieves path tracing performance comparable to strong closed-source models, and yields further gains when combined with label supervision.

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.

Problem

• spatial reasoning
• imaginative perception
• vision language models
• occluded spaces
• unobserved spatial structure

Innovation

• Imaginative Perception Tokens
• Spatial Reasoning
• Vision Language Models
• Multimodal Learning
• Intermediate Representations