AI models that can ace static spatial questions fall apart the moment they have to explore a space themselves — revealing a fundamental gap between knowing and discovering that could redefine how we build navigating AI.
A new benchmark called "Theory of Space" was created to test whether AI can actively explore unknown environments and build internal spatial maps — going beyond existing benchmarks that only test passive observation or goal completion, without checking if the AI truly understands the space.
Active exploration is dramatically harder than passive observation for AI. When models had to explore on their own (deciding where to move and look), accuracy dropped significantly — for example, GPT-5.2 fell from 57.1% to 46.0% and Gemini-3 Pro from 60.5% to 57.3% in vision-based environments. Think of the difference between being handed a complete map vs. having to draw one yourself while walking through a building.
AI explores inefficiently. Simple scripted programs reached adequate coverage of a space in ~9 steps; state-of-the-art AI models needed 14+ steps and still built less accurate maps — often rushing toward every new doorway while leaving rooms only half-explored.
Vision is a major weak point compared to text. Across all tests, models performed significantly better when given text descriptions of their surroundings (e.g., "chair is front-left and near") than when given actual images. This reveals that visual perception errors — not just spatial reasoning — are holding AI back.
AI is especially bad at tracking object orientation in visual environments. When using images, models correctly identified object orientations only 20–32% of the time, compared to 91–92% when using text descriptions — a massive gap for something humans do almost effortlessly.
Spatial memory degrades over time. As AI agents explored further, their beliefs about previously seen objects became less accurate (stability scores of only 56–62% in vision settings), meaning new observations were corrupting earlier correct memories — like overwriting good notes with bad ones.
AI suffers from "belief inertia" — it struggles to update its mental map when things change. When objects were secretly moved after initial exploration (a "false belief" test borrowed from child psychology), agents frequently failed to overwrite their old spatial memory. GPT-5.2's orientation inertia reached 68.9% in vision settings, meaning it kept "remembering" objects in the wrong place even after directly seeing them in new locations.
The benchmark externalizes AI's internal spatial understanding by requiring models to output structured "cognitive maps" at each step (formatted as JSON — a structured data format), transforming evaluation from "did you get the right answer?" to "do you actually understand the space?" — making the AI's reasoning process visible and testable.
Navigation and spatial awareness are foundational to almost everything we want AI to do in the physical world — from robots that assist in hospitals or warehouses, to autonomous vehicles, to AI assistants that help you find your way around a new building.
This research reveals that today's most powerful AI models have a critical blind spot: they can reason about space when handed complete information, but they struggle to discover spatial structure on their own. That's the difference between a GPS that works when it has a map and one that can build the map from scratch.
The "belief inertia" finding is particularly important — an AI that can't update its mental model when the world changes is unreliable in any real environment, where things move, rooms get rearranged, and situations evolve. By making these failures measurable and visible, this benchmark gives researchers a clear target for building AI that can truly navigate, adapt, and understand the physical world — not just answer questions about it.
The researchers built their benchmark from the ground up, designing it to test something no existing benchmark measured: whether an AI can actively discover the structure of an unknown space, not just answer questions about a space it has already been shown.
The framework rests on an analogy to a well-known concept from child psychology called Theory of Mind — the ability to understand what another person believes, even when that belief differs from reality. For example, a child who understands Theory of Mind knows that if you hide a toy while their friend is out of the room, the friend will still think the toy is in its original spot. The researchers borrowed this logic and applied it to physical space: just as Theory of Mind tests whether you can model another person's hidden mental state, Theory of Space tests whether an AI can model the hidden physical structure of an environment it has never seen.
This framework breaks down into three specific, testable abilities:
To test these three abilities, the researchers created procedurally generated multi-room environments — meaning the room layouts were automatically produced by a program, generating many different configurations rather than relying on a single hand-crafted scenario. This ensures results are not specific to one particular layout.
They ran experiments in two versions of each environment:
Running both versions side-by-side allowed the researchers to separate two different sources of failure: spatial reasoning (can the AI think about space?) versus visual perception (can the AI correctly interpret what it sees?).
The exploration task itself works like this: an AI agent starts in an unknown environment with no map. At each step, it must choose an action — rotate left or right, move forward, enter a new room — and after each action, it receives either a text description or an image of what it currently sees. Crucially, after each step, the agent must output a cognitive map formatted as JSON (a structured data format, like a detailed checklist with labels and values) listing every object it has observed, along with each object's estimated position and orientation. This happens continuously throughout exploration, not just at the end.
Think of it like asking someone to hand you their notes after every single step of a tour, rather than just asking them to describe the building when the tour is over. This makes the AI's internal understanding visible and testable at every moment.
The false belief test is a particularly revealing case study. After an agent finishes its initial exploration and has built a map, the researchers secretly move objects to new locations. The agent then re-explores the same environment. The question is: will the agent notice the changes and update its map, or will it stubbornly hold onto its original (now incorrect) beliefs?
This is directly borrowed from a classic child psychology experiment. In the original test, a child watches someone place a toy in a box, then leave the room. While the person is gone, the toy is moved to a different box. When asked where the person will look for the toy, young children (under about age 4) say the new location — they cannot separate what they know from what the other person knows. The AI equivalent is the reverse problem: the AI has seen the new location, but still reports the old one — it cannot separate its old memory from its new observation.
The scripted proxy agent serves as a useful comparison point. The researchers built a simple, rule-based program (not an AI — just a set of fixed instructions) that systematically covers a space. This scripted agent reached adequate coverage in approximately 9 steps. The state-of-the-art AI models needed 14 or more steps and still produced less accurate maps. The scripted agent has no intelligence, but it has a clear, consistent strategy — explore methodically, do not revisit what you have already seen. The AI models, by contrast, frequently moved toward every new doorway immediately, leaving rooms only partially examined before moving on.
Most existing spatial benchmarks fall into one of two categories:
Theory of Space is different in two important ways:
First, it requires active information gathering. The AI must decide what to look at next, which means exploration strategy is part of the test. A model that explores poorly will build a poor map, even if its reasoning ability is strong.
Second, it makes the AI's internal understanding directly measurable. By requiring a structured cognitive map output at every step, the researchers can evaluate not just whether the AI got the final answer right, but whether its internal model of the space was accurate, stable, and correctly updated over time. This is the difference between grading only the final exam answer versus also reading the student's notes throughout the course.
The four-part evaluation framework — correctness, perception, stability, and self-tracking — is also novel. Most benchmarks measure only one thing (final answer accuracy). Here, each dimension reveals a different type of failure. For example, a model might have high correctness early in exploration but low stability later, meaning it starts well but corrupts its own good memories as it gathers more information.
Challenge 1: How do you test something invisible?
The core difficulty is that an AI's internal spatial understanding is not directly observable. You can ask it questions, but a correct answer does not prove genuine understanding — the model might be guessing or using shortcuts. The researchers solved this by requiring the AI to externalize its beliefs as a structured cognitive map at every step. This transforms an invisible internal state into a measurable, comparable output. The JSON format was chosen because it is structured and unambiguous — there is no room for vague or partially correct answers about object positions.
Challenge 2: Separating perception failures from reasoning failures
If an AI performs poorly in a vision-based environment, is it because it cannot reason about space, or because it cannot correctly interpret the images? These two problems require very different solutions. The researchers addressed this by running every experiment in both text and vision conditions. The results were striking: orientation identification accuracy was 97.9% in text but only 14.3% in vision for GPT-5.2. This clearly shows that visual perception — not spatial reasoning — is the primary bottleneck in vision settings, giving researchers a precise target for improvement.
Challenge 3: Measuring belief revision, not just belief construction
Testing whether an AI can update a map is harder than testing whether it can build one, because you need a baseline to compare against. The false belief paradigm solved this cleanly: establish a known initial state, change it in a controlled way, then measure how accurately the agent's map reflects the new state versus the old one. The belief inertia metric — measuring how often an agent keeps reporting an object's old position even after directly observing it in a new location — gives a precise, quantifiable measure of this failure. GPT-5.2's orientation inertia of 68.9% in vision settings means that in roughly 7 out of 10 cases where the agent directly saw an object had moved, it still reported the original location.
Challenge 4: Ensuring results generalize across many environments
A benchmark tested on only one or two room layouts could produce results that are specific to those layouts rather than revealing general AI capabilities. Procedural generation — automatically creating many different room configurations — addresses this by ensuring the benchmark tests general spatial ability rather than memorization of a specific environment.