Artificial intelligence models are gaining a new superpower: the ability to mentally visualize objects from different angles during reasoning tasks. A new research framework called "Thinking with Novel Views" (TwNV) shows that when AI models generate alternative camera perspectives while solving spatial problems, they achieve significantly better accuracy than models that rely on a single static image.

Why Can't AI Models Understand Space From One Image?

Current large multimodal models (LMMs), which process both text and images, struggle with spatial reasoning tasks because they're confined to a single viewpoint. When asked to understand 3D relationships, object orientations, or how multiple objects relate to each other in space, these models hit a wall. The problem isn't a lack of intelligence; it's a fundamental limitation of perspective. A single image can hide crucial information. Occluded geometry, hidden relationships, and spatial ambiguities remain invisible from one angle.

Researchers tested whether simple 2D image manipulations like cropping or zooming could help. The results were disappointing. When GPT-5 was equipped with tools to identify and crop task-relevant regions, overall spatial reasoning accuracy actually dropped by 0.8 percentage points, and performance on multi-object relationship questions fell by 2.0 percentage points. Cropping and zooming cannot reveal what's hidden or clarify 3D structures that simply don't appear in the original view.

How Does the Novel Views Approach Work?

The TwNV framework operates as a three-stage pipeline that mirrors how humans mentally solve spatial puzzles. First, an AI model acts as a "Planner," identifying which alternative camera angle would best clarify the spatial question. Second, a generative image model serves as a "Synthesizer," rendering the requested viewpoint by simulating camera movements like translation, panning, or tilting. Finally, the AI model acts as a "Reasoner," jointly interpreting both the original and synthesized images to produce an answer.

This approach transforms a static single-image query into an iterative, generation-augmented reasoning process. Rather than treating image generation as a creative end-goal, researchers reconceptualized it as a dynamic, 3D-aware reasoning workspace. The results are consistent: across four different AI models spanning both closed-source systems and open-source alternatives, TwNV improved accuracy by 1.3 to 3.9 percentage points, with the largest gains on viewpoint-sensitive tasks like orientation and multi-object relationships.

What Makes This Approach Different From Previous Methods?

Earlier attempts to inject 3D understanding into AI models relied on depth maps, point clouds, or explicit 3D reconstruction. However, recovering 3D geometry from a single image is fundamentally unreliable and scale-ambiguous. The geometry these methods inject is intrinsically noisy. Additionally, they discard the image priors, textures, and semantic context that pretrained AI models are trained to understand.

Generative novel-view synthesis avoids both trade-offs. Pretrained image generators have absorbed 3D-aware, world-model-level visual priors during large-scale web training and can be repurposed to render the same scene from a different camera pose. This produces a new image that keeps the AI model within its trained input distribution while supplying geometric evidence the original view lacks.

How to Optimize Novel-View Generation for Better Results

• Use Precise Camera Parameters: Continuous 3D camera parameters, such as explicit translation offsets and rotation angles, vastly outperform both free-form natural language descriptions and discrete categorical instructions. Providing explicit geometric priors minimizes semantic noise and ensures the generation of geometrically precise viewpoints.
• Prioritize Generation Fidelity: The quality of the synthesized view is tightly coupled with downstream spatial accuracy. A specialized pose-aware novel-view editing model ensures geometric precision, and better synthesis directly yields better reasoning outcomes.
• Implement Iterative Refinement: Inference-time visual scaling through iterative multi-turn view refinement further improves performance. In each round, the Reasoner evaluates its previous generation, the Planner updates camera instructions, and the Synthesizer produces improved viewpoints, allowing the system to progressively converge on the most diagnostic visual evidence.

Which AI Models Benefit Most From This Approach?

The TwNV paradigm delivers consistent improvements across a diverse spectrum of model capacities. Testing included frontier closed-source systems like Gemini-3-Flash and GPT-5, as well as open-source models like the Qwen3-VL series with 235 billion and 32 billion parameters. Notably, smaller models showed a "Small-Model Dividend," with relative gains more pronounced in parameter-constrained systems. For example, Qwen3-VL-32B saw a 3.9 percentage point improvement, compared to 1.3 percentage points for GPT-5. This suggests that explicit view synthesis serves as a compensatory mechanism for smaller models by offloading 3D reasoning to an external visual workspace.

Performance gains are non-uniform across spatial subtasks. Viewpoint-sensitive categories like orientation and multi-object relationships benefit the most, while size estimation sees slight degradation, likely because novel views alter apparent object scale. This specificity highlights that the approach is not a universal fix but rather a targeted enhancement for particular types of spatial reasoning.

What Does This Mean for the Future of AI Reasoning?

The TwNV framework echoes recent scaling trends in language reasoning, where models improve by thinking longer and exploring more possibilities. This research establishes novel-view generation as a practical lever for advancing spatial intelligence in multimodal AI systems. As AI models become more capable, the ability to dynamically generate and reason about alternative perspectives could unlock new applications in robotics, autonomous vehicles, 3D design, and scientific visualization.

The findings suggest that test-time compute, the computational resources spent during inference rather than training, is becoming an increasingly important frontier. By allocating more compute to reasoning processes at test time, models can achieve better results without requiring larger parameter counts or more expensive training procedures. This shift has profound implications for how AI systems are designed and deployed in the future.