When and How Much to Imagine:
Adaptive Test-Time Scaling with World Models
for Visual Spatial Reasoning

Despite rapid progress in Multimodal Large Lan- guage Models (MLLMs), visual spatial reasoning remains unreliable when correct answers depend on how a scene would appear under unseen or alternative viewpoints. Recent work addresses this by augmenting models with world models for visual imagination, yet when such imagina- tion is actually necessary, how much is benefi- cial, and when it becomes harmful, remain poorly understood. In practice, indiscriminate imagina- tion can increase computation and even degrade performance by introducing misleading evidence. In this work, we present an in-depth analysis of test-time visual imagination as a controllable re- source for spatial reasoning. We study when static visual evidence is sufficient, when imagination improves reasoning, and how excessive or unnec- essary imagination affects accuracy and efficiency. To support this analysis, we introduce AVIC, an adaptive test-time framework with world mod- els that explicitly reason about the sufficiency of current visual evidence before selectively invok- ing and scaling visual imagination. Across spa- tial reasoning benchmarks (SAT, MMSI) and an embodied navigation benchmark (R2R), our re- sults reveal clear regimes where imagination is critical, marginal, or detrimental, and show that selective control can match or outperform fixed imagination strategies with substantially fewer world-model calls and language tokens. Together, our findings highlight the importance of analyzing and controlling test-time visual imagination for efficient and reliable spatial reasoning.

Figure 2: Comparison between our Adaptive Visual Imagination Control (AVIC) and other methods. (a) Answers directly from the current observation without any imagination. (b) Always invokes the world model w. full exploration to generate imagined views for downstream reasoning. (c) Ours: Uses a policy model to first decide whether visual imagination is necessary and to plan actions accordingly. It selectively queries the world model (both when and how much) and otherwise performs direct reasoning.

Table 1: Comparison between test-time scaling methods on SAT-Real dataset with different MLLMs. The best results are denoted by the bold. Avg. WM: average world model calling times over the dataset.

We evaluate both reasoning performance and computational cost, measured by total token usage and average world model calling times. Across all open-source and proprietary MLLM models, our method consistently improves average accuracy over the corresponding base MLLMs and achieves performance competitive with, or superior to, the always-on/dense imagination baseline (MindJourney). With GPT-4.1, our method raises average accuracy from 74.0% to 79.3%, surpassing MindJourney while using far fewer tokens. With o1, we achieve the best overall accuracy (85.3%), improving by 10.7% over the base model.

Table 2: Evaluation results on MMSI-Bench. Best results are highlighted in bold.

we further apply our method to another visual spatial reasoning benchmark, MMSI-Bench, and observe consistent improvements across MLLMs, highlighting the generalizability of our approach.

Table 3: Results on R2R embodied navigation dataset.

We further applied our adaptive visual test-time scaling method to the embodiment navigation task, and compared it with previous works.in MapGPT, and re-implement MapGPT with the same API. We apply our method within MapGPT's step-wise navigation framework. Compared to MapGPT with GPT-4o, our method achieves higher OSR, SR, and SPL while also reducing navigation error (NE), indicating more reliable goal reaching with shorter, less redundant trajectories. These gains suggest that our world-model imagination strategy helps resolve ambiguous visual-spatial decisions. Overall, the results demonstrate that our adaptive visual test-time scaling is an effective mechanism and transfers the gains for embodied navigation.

Across our analysis, we found:
(1) WM should be used selectively, primarily when spatial reasoning requires predicting future states under hypothetical actions, rather than reinterpreting existing visual evidence.
(2) Visual spatial reasoning benefits most from targeted rather than extensive WM imagination.

Qualitative examples on SAT of the always-on imagination method and our adaptive method, as well as the R2R navigation task. In the navigation example, the green option is selected by the model with adaptive imagination via our method, while the red one is without world model imagination. We compare our adaptive visual TTS method with the always-on imagination method, MindJourney (MJ). In the first example, the target object (the cash counter) is already clearly visible in the observed view. Our method correctly identifies that additional visual imagination is unnecessary and directly skips world model. In contrast, MJ indiscriminately invokes the world model, generating multiple imagined views that introduce misleading evidence and ultimately lead to an incorrect prediction. In the second example, AVIC yields the correct answer by selectively imagining the state where the agent is in front of the trash bin. In contrast, MJ performs dense imagination and generates views that do not accurately reflect this critical spatial condition, leading to an incorrect prediction. Furthermore, we present a qualitative navigation example at the bottom. Our adaptive visual test-time scaling selectively augments informative indoor observations (e.g., zooming in or turning to explore nearby views), enabling the agent to better inspect the environment and align its actions with the global instruction (“go to the kitchen”). In contrast, the baseline without visual imagination lacks sufficient perceptual evidence and consequently chooses an incorrect direction.