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

Abstract: Despite rapid progress in Multimodal LLMs (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 reasoning with world models for visual imagination, but questions such as when imagination is actually necessary, how much of it is beneficial, and when it becomes harmful, remain poorly understood. In practice, indiscriminate imagination 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 resource for spatial reasoning. We study when static visual evidence is sufficient, when imagination improves reasoning, and how excessive or unnecessary imagination affects accuracy and efficiency. To support this analysis, we introduce AVIC, an adaptive test-time framework with world models that explicitly reasons about the sufficiency of current visual evidence before selectively invoking and scaling visual imagination. Across spatial reasoning benchmarks (SAT, MMSI) and an embodied navigation benchmark (R2R), our results reveal clear scenarios 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. Overall, our findings highlight the importance of analyzing and controlling test-time imagination for efficient and reliable spatial reasoning.

• The paper presents AVIC, a framework that adaptively scales world model usage to enhance spatial reasoning while conserving computation.
• It employs policy-guided gating and targeted action planning to counteract errors from excessive or redundant visual imagination.
• Empirical results show significant accuracy improvements and efficiency gains on SAT, MMSI, and navigation benchmarks using adaptive control.

Adaptive Visual Imagination Control: Instance-Level Test-Time Scaling for Spatial Reasoning

Motivation and Empirical Analysis of Always-On Visual Imagination

The paper addresses a central issue in visual spatial reasoning with Multimodal LLMs (MLLMs): determining when and how much to invoke world-model-based visual imagination. Spatial queries often require information not directly accessible from a static egocentric observation, making generative imagination via world models essential. However, previous efforts frequently adopted indiscriminate, exhaustive invocation strategies, ignoring the variable necessity and efficacy of imagination across instances.

A structured empirical analysis reveals three primary regimes for always-on imagination: (1) helpful—imagination reveals unseen viewpoints and improves reasoning; (2) misleading—imagination introduces errors due to degraded visual content or absent task-relevant objects; (3) unnecessary—required information is observable in the original view, rendering imagined views redundant. The majority of cases are unnecessary, with only a minority benefiting from imagination (Figure 1).

Figure 1: Three case regimes of always-on visual imagination: helpful, misleading, and unnecessary.

Furthermore, statistical evaluation demonstrates that excessive imagination incurs substantial computational overhead with limited accuracy gains. Increasing the number of imagined views does not monotonically improve performance; accuracy saturates and even degrades when too many views are generated. This inefficiency motivates adaptive control, as uniform world-model invocation neither optimally leverages computational resources nor improves reliability.

Figure 2: Distribution of always-on imagination cases, highlighting unnecessary invocation as dominant.

Framework: AVIC—Adaptive Visual Imagination Control

The Adaptive Visual Imagination Control (AVIC) framework introduces instance-wise, policy-guided regulation to test-time world-model invocation. For each spatial reasoning query, the policy model first decides whether additional visual evidence is necessary. If not, the answer is inferred directly from the observed image(s). Otherwise, the policy generates a conditional action plan specifying explicit viewpoint transformations or navigational steps, which are rendered by the world model.

Key design elements:

• Policy-Guided Gating: A model explicitly performs uncertainty-aware deliberation on sufficiency of current visual evidence, reducing arbitrary computation.
• Sampled Policy Self-Consistency: Multiple policy samples are aggregated via majority voting, improving decision robustness.
• Targeted Action Planning: Rather than exhaustive trajectory expansion, short, discrete action plans focus imagination on informative viewpoints.
• Trajectory-Level Verification: Instead of scoring isolated imagined keyframes, entire action-sequence trajectories are evaluated for temporal and spatial consistency, mitigating inter-frame information loss.
• Instance-Dependent Scaling: AVIC scales imagination adaptively, maintaining computational parsimony while achieving high accuracy.

Framework comparisons illustrate that AVIC, by explicit policy gating and action scaling, outperforms fixed baselines with fewer world model calls and lower computational cost (Figure 3).

Figure 3: Methodological comparison: direct QA, always-on imagination, and AVIC's selective invocation.

Empirical Results and Ablation Analyses

Evaluation across spatial reasoning (SAT, MMSI) and embodied navigation (R2R) benchmarks confirms that AVIC achieves state-of-the-art or competitive results, consistently improving accuracy versus baselines. Notable numerical outcomes include:

• On SAT-Real, AVIC improves average accuracy up to 79.3% with GPT-4.1 (compared to 74.0% vanilla and 77.3% always-on), while reducing token usage and runtime by an order of magnitude.
• AVIC typically invokes the world model ≤1 times per instance, with accuracy gains saturating as more imagined views are generated.

Ablation studies show that binary gating alone, without action-level scaling, reduces computation but underperforms due to suppressed imagination. Combining gating with adaptive action scaling yields optimal performance and efficiency, underscoring the necessity of both selective invocation and targeted trajectory generation.

Task-level error analysis identifies four primary error types: limited observability, viewpoint dependence, action-conditioned reasoning, and dynamics understanding. The utility of world models is highest for action-conditioned spatial reasoning (accuracy improvement +57.1%), as these queries inherently require prediction of post-action visual states. For static viewpoint transformations or object motion, world models are less critical (Figure 4).

Figure 4: Error-type distributions across task categories, highlighting compositional structures in spatial reasoning failures.

Qualitative Examples and Generalization

Qualitative analysis reveals that always-on methods may introduce misleading or redundant evidence, resulting in incorrect predictions. AVIC, by skipping unnecessary imagination and selectively simulating critical spatial states, avoids distractors and aligns action plans with the task rationale. In navigation tasks, AVIC enhances exploration and goal alignment by augmenting perceptual evidence only at ambiguous decision points (Figure 5).

Figure 5: Qualitative comparison showing AVIC's selective augmentation and navigation improvements.

The approach generalizes across spatial reasoning settings, providing transferable gains from static QA tasks to embodied navigation. Policy and QA model selection interacts strongly with performance: a stronger QA model amplifies the benefits, but policy design critically governs efficiency.

Practical and Theoretical Implications

Practically, AVIC recommends minimizing test-time world-model invocation to scenarios with actionable spatial ambiguity or counterfactual queries. The framework aligns resource allocation with task structure, reducing unnecessary computation and improving reasoning reliability. Theoretically, the findings suggest that spatial reasoning with MLLMs can be decomposed into error-aware invocation policies and informed trajectory sampling, rather than brute-force simulation. This approach paves the way for further research on state-aware, error-driven adaptive computation in multimodal reasoning agents.

For future development, integrating finer-grained error detection, dynamic policy refinement, and feedback-driven imagination may further enhance spatial reasoning capability. The paradigm can extend to other modalities, including dynamic scene simulation, interactive visual planning, and RL-based embodied agents.

Conclusion

Adaptive control of visual imagination, as implemented in AVIC, is essential for efficient and reliable spatial reasoning with world models. Indiscriminate always-on strategies are computationally wasteful and often detrimental in accuracy. AVIC achieves selective, instance-level scaling by policy-guided gating and targeted action planning, yielding strong empirical results across benchmarks and task categories. The work advances both practical deployment and theoretical understanding of adaptive computation for multimodal spatial intelligence (2602.08236).