Show, Don’t Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text | Xu Wang et al. | ResearchPod

Transcript: Show, Don’t Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text

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.

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.

Show, Don’t Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text

Xu Wang, Kaixiang Yao, Miao Pan, Xiaohe Zhou, Xuanyu Liu, Wenqi Zhang, Xuhong Zhang

Jul 23, 2026·5:18·arxiv·Original Paper

Spatial IntelligenceProVisESpatialGen-BenchPixel-space ExpressionInterface MismatchAnswer-Interface Mismatch

Listen

5 min

Briefing

Abstract

Spatial intelligence is essential for agents to move from static semantic understanding toward interacting with the physical world. Many spatial tasks are grounded in continuous visual scenes, where locations, regions, and paths are more naturally expressed by pointing, marking, or drawing than by reporting precise coordinates or discrete textual symbols. Yet existing spatial reasoning benchmarks usually require coordinates, options, or text, creating an answer-interface mismatch for image-generation models. This makes it difficult to evaluate image-generation models under the same task semantics as text-output VLMs, despite their ability to externalize spatial judgments directly in pixel space. We propose ProVisE (Protocolized Visual Evaluation), a benchmark-agnostic framework that elicits protocol-constrained visual answers from image-generation models and parses them into structured predictions compatible with original metrics. ProVisE also includes an Agentic builder that constructs and validates task-specific protocols for new benchmarks. We further introduce SpatialGen-Bench, a curated diagnostic benchmark of 470 samples across 14 spatial subtasks, four capability levels, and diverse answer forms. We evaluate representative text-output VLMs and image-generation models in a unified setting and validate Agentic protocol construction on six external spatial benchmarks. Results show that image-generation models are competitive when spatial answers can be externalized directly in pixel space, while text-output VLMs retain a clear advantage in compositional spatial reasoning. These findings reveal complementary strengths of pixel-space expression and text-based reasoning and establish a metric-compatible testbed for studying spatial cognition in image-generation models.

Research Context and Motivation

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 ProVisE Framework

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.

SpatialGen-Bench and Key Findings

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.

Key Terms and Definitions

ProVisE — A framework that enables the evaluation of image-generation models on spatial benchmarks by converting visual outputs into structured, metric-compatible predictions.
Answer-Interface Mismatch — The discrepancy between the natural visual output of image-generation models and the textual or coordinate-based requirements of traditional spatial benchmarks.
Agentic Builder — An automated component of ProVisE that constructs and validates generation-parser protocols for new benchmarks by analyzing task contracts.
SpatialGen-Bench — A diagnostic benchmark comprising 14 spatial subtasks categorized by capability level, designed to evaluate spatial cognition across diverse answer formats.
Protocol-Constrained Visual Answer — A model-generated image that is restricted to a specific visual format (e.g., a mask or point) to ensure it can be reliably parsed into a structured answer.

Show, Don’t Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text: Research Paper Podcast