Title: Optimizing Test-Time Scaling: Adaptive Imagination via World Models for Visual Spatial Reasoning

Abstract:

Although Multimodal Large Language Models (MLLMs) have advanced rapidly, their visual spatial reasoning capabilities remain inconsistent, particularly when determining how a scene appears from unseen or alternative perspectives. While recent studies have attempted to bolster reasoning through world-model-based visual imagination, critical questions persist regarding the precise timing, optimal quantity, and potential downsides of such imaginative processes. In real-world applications, unselective imagination can escalate computational costs and impair performance by introducing confounding evidence.

This study offers a comprehensive examination of test-time visual imagination as a manageable resource for spatial reasoning. We investigate the conditions under which static visual data is adequate, identify scenarios where imagination enhances reasoning, and analyze the impact of superfluous or excessive imagination on both accuracy and efficiency. To facilitate this analysis, we propose AVIC, an adaptive framework that leverages world models to explicitly evaluate the sufficiency of current visual evidence before deciding whether to engage visual imagination.

Furthermore, to acquire this gating and planning behavior without requiring annotations on the necessity or extent of imagination, we introduce AVIC-R. This variant trains the policy using GRPO, deriving rewards from question-answer correctness and penalties based on the cost of imagination. Our evaluation across spatial reasoning benchmarks (SAT, MMSI) and an embodied navigation benchmark (R2R) delineates specific contexts where imagination is essential, offers limited benefit, or proves harmful. The results demonstrate that selective control mechanisms can achieve performance levels equal to or superior to fixed imagination strategies, while significantly reducing the number of world-model calls and language tokens. Notably, AVIC-R outperforms robust proprietary baselines, such as GPT-4o and GPT-4.1, despite utilizing the world model less frequently. These findings underscore the necessity of analyzing and regulating test-time imagination to ensure efficient and dependable spatial reasoning.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC