Author-updated Summary
Verified author edit
Spatial intelligence is a critical requirement for AI agents to interact with the physical world. While current research primarily evaluates spatial reasoning through text-output vision-language models (VLMs), this approach creates an 'answer-interface mismatch.' Many spatial tasks—such as pointing, marking regions, or drawing paths—are naturally continuous and visual. Forcing these tasks into textual or coordinate-based outputs is unnatural for both the model and the task, and it prevents image-generation models from being evaluated on equal footing with text-output models.
The authors introduce ProVisE (Protocolized Visual Evaluation), a benchmark-agnostic framework designed to bridge this gap. ProVisE allows image-generation models to provide visual answers that are then parsed into structured predictions compatible with original benchmark metrics. The framework includes an 'Agentic builder' that automatically constructs and validates task-specific protocols, ensuring that the evaluation remains consistent across different benchmarks without requiring manual engineering for every new task. This allows for a direct, metric-compatible comparison between text-output VLMs and image-generation models under shared task semantics.
To test this framework, the authors developed SpatialGen-Bench, a diagnostic suite containing 470 samples across 14 spatial subtasks. These tasks are organized into a hierarchy of four capability levels: perception, understanding, reasoning, and interaction. By evaluating representative models, the study reveals that image-generation models perform strongly when spatial evidence can be externalized as pixels (e.g., marking regions or paths). Conversely, text-output VLMs remain superior in tasks requiring complex, compositional spatial reasoning. These results highlight that the two model architectures offer complementary strengths, suggesting that future spatial intelligence research should leverage both pixel-space expression and text-based reasoning.
Alex: Welcome to another episode of ResearchPod. Today, we're looking at a new way to measure how well AI models understand physical space.
Sam: Is this paper arguing that we've been testing AI's spatial intelligence the wrong way?
Alex: That's essentially the argument, yes. Think about what it means to understand space. If you asked someone to prove they knew where a chair was in a room, the natural thing to do is point at it—not recite its coordinates. But right now, we test AI spatial reasoning almost entirely by asking models to produce text or numbers. For AI systems that work visually—that "think" in images—that's a bit like grading an artist by asking them to write an essay about their painting instead of just looking at the painting.
Sam: So the mismatch isn't just inconvenient—it might actually be hiding how capable these visual models are.
Alex: That's the core concern. When you force a visual model to translate its understanding into words, you're adding an extra step that can introduce errors. The model might understand the space perfectly but stumble when converting that understanding into text. The paper argues this makes it genuinely difficult to know what these models can and can't do.
Sam: So what's the proposed fix?
Alex: The research introduces a framework called ProVisE—short for Protocolized Visual Evaluation. The idea is to let visual models answer in their native language, which is images, rather than forcing them to speak in text. Instead of asking a model "where is the cup?"—and waiting for a text answer—ProVisE asks it to draw a shape over the cup in the image. That drawing is the answer.
Sam: But how does a drawing become a grade? You can't just compare two images and say one is right or wrong.
Alex: Right, and that's the considered part. ProVisE uses what it calls a "visual protocol"—a strict set of rules about how the model must draw its answer. It might require the model to shade a specific region, or draw a bounding box in a particular color. Because the rules are precise, a piece of software called a parser can read the drawing and convert it into a standard coordinate or label—something a computer can score automatically, the same way it would score a multiple-choice test.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.
Sam: So it's like a fill-in-the-blank test, except instead of writing a word in the blank, the model draws a shape. And because the shape has to follow strict rules, the computer can read it.
Alex: That's exactly it. The system also includes what the paper calls an "Agentic builder"—an automated tool that constructs these drawing rules for new types of tasks, so researchers don't have to manually design a protocol every time they want to test something new.
Sam: And once you have this, you can finally compare visual models and text-based models on equal footing?
Alex: That's the goal. By converting both kinds of output—drawings and text—into the same standardised format, you can score them using identical metrics. The playing field becomes level in a way it hasn't been before.
Sam: So when you actually run the comparison, what does the paper find?
Alex: The finding is more nuanced than a simple winner. The paper describes complementary strengths. When the task involves marking or identifying regions in an image—something naturally expressed in visual space—image-generation models perform quite competitively. They're doing what they were built to do.
Sam: But I'd guess text-based models still have an edge in certain areas?
Alex: They do. Where text-based models remain stronger is in what the paper calls compositional reasoning. Think of tasks that require several logical steps in sequence—like "find the object to the left of the red box, but not touching the wall." That kind of multi-step spatial logic is harder to express by drawing, and text-based models handle it more reliably.
Sam: So it's less about one being superior and more about matching the right tool to the right kind of question.
Alex: That's a fair summary. And there's a further insight that comes out of the evaluation. When visual models fail on these tests, they almost always still produce a valid, parseable drawing—an image that follows the protocol correctly, but gives the wrong answer. The errors are logical, not mechanical.
Sam: That's a meaningful distinction. It means the drawing interface itself isn't the problem—the model understood the instructions for how to answer, it just reasoned incorrectly about what the answer should be.
Alex: Exactly. The paper describes this as clearing what you might call "interface fog." Once you remove the friction of forcing visual models to speak in text, you can see their actual reasoning capabilities more clearly. The failures you observe are genuine reasoning gaps, not artefacts of an awkward translation process.
Sam: Which makes the benchmark more honest. You're measuring what the model actually knows, not how well it can describe what it knows.
Alex: That's the underlying motivation for the whole framework. ProVisE, paired with the new benchmark the paper introduces—SpatialGen-Bench—gives researchers a testbed where visual models are evaluated on their own terms. It's a step toward understanding what these systems genuinely understand about the physical world, rather than what they can say about it. And that distinction, it turns out, matters quite a lot.
Sam: It makes you think about how many other areas of AI evaluation might have the same problem—where the way we're asking the question is getting in the way of seeing the real answer.
Alex: That's a question the paper raises implicitly, and it's worth sitting with. The methodology here is specific to spatial reasoning, but the principle—that the output format of a test should match the native output format of the system being tested—seems like it could apply more broadly. Thanks for listening to ResearchPod.