# How and What to Imagine? Visual Thinking in Unified Multimodal Models for Cross-View Spatial Reasoning

Source: [How and What to Imagine? Visual Thinking in Unified Multimodal Models for Cross-View Spatial Reasoning](<https://www.sota2.com/research/paper/6a167ad9dbd7150506f802fe>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-31T06:00:01.542Z

Authors: Qian Yang, Ankur Sikarwar, Huy Le, Le Zhang, Zhuan Shi, Perouz Taslakian, Aishwarya Agrawal

Published: 2026-05-26T17:20:05.000Z

## Abstract

Cross-view spatial reasoning remains a weak spot for vision-language models (VLMs): they often reason in language and lose the fine-grained geometry needed for the task. Thinking with images aims to address this by generating an intermediate thinking image, but recent work shows that models often ignore the visual evidence in these traces. We therefore ask how to make visual thinking matter, and what kind of visual thinking works best. We study these questions in unified multimodal models (UMMs), which natively support interleaved image-text generation. For the first question, we propose View Dropout (VDrop), a training-time intervention that hides parts of one input view from the answer span while keeping them visible to the thinking-image tokens. This encourages the model to use the thinking image when answering, instead of relying only on the input views. Once the thinking image is used for answer prediction, we study which type of visual thinking is most effective. We frame this as a learnability-informativeness tradeoff and compare three thinking-image variants: top-down, panoramic, and point-matching renderings. Trained on synthetic scenes and evaluated on five real-world out-of-domain benchmarks, panoramic visual thinking with VDrop is the only configuration that is both informative and learnable, and it achieves the best out-of-domain generalization.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Cross-view spatial reasoning on COSMIC in-domain](<https://www.sota2.com/research/sota/cross-view-spatial-reasoning-on-cosmic-in-domain.md>): Cross-view spatial reasoning; COSMIC in-domain; Anchor: 94.4.

- [OOD Generalization on Real-world Cross-View Spatial Reasoning Benchmarks (Aggregate)](<https://www.sota2.com/research/sota/ood-generalization-on-real-world-cross-view-spatial-reasoning-benchmarks-aggregate.md>): Generalization across real-world cross-view spatial reasoning benchmarks; OOD Evaluation Suite Aggregate; Avg OOD Performance: 40.

- [Integrating evidence across multiple images on BLINK MultiView](<https://www.sota2.com/research/sota/integrating-evidence-across-multiple-images-on-blink-multiview.md>): Integrating evidence across multiple images; BLINK MultiView; MultiView Score: 62.4.

- [Real-world cross-view spatial reasoning on MMSI](<https://www.sota2.com/research/sota/real-world-cross-view-spatial-reasoning-on-mmsi.md>): Real-world cross-view spatial reasoning; MMSI; Overall Accuracy: 32.

- [Real-world cross-view spatial reasoning on OmniSpatial](<https://www.sota2.com/research/sota/real-world-cross-view-spatial-reasoning-on-omnispatial.md>): Real-world cross-view spatial reasoning; OmniSpatial; CL Score: 38.9.

- [Real-world cross-view spatial reasoning on STARE](<https://www.sota2.com/research/sota/real-world-cross-view-spatial-reasoning-on-stare.md>): Real-world cross-view spatial reasoning; STARE; Accuracy: 35.6.

- [Multi-view spatial reasoning on MindCube](<https://www.sota2.com/research/sota/multi-view-spatial-reasoning-on-mindcube.md>): Multi-view spatial reasoning; MindCube; Accuracy: 36.5.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2605.27310>)

[PDF](<https://arxiv.org/pdf/2605.27310.pdf>)